Provided by the author(s) and University College Dublin Library in accordance with publisher
policies. Please cite the published version when available.
Title
Functional literacy, educational attainment and earnings : a multi-country comparison
Authors(s)
Denny, Kevin; Harmon, Colm; O'Sullivan, Vincent
Publication date
2003-08
Series
UCD Centre for Economic Research Working Paper Series; WP03/19; ISSC Discussion
Paper Series; 2003/12
Publisher
University College Dublin. School of Economics; University College Dublin. Institute for the
Study of Social Change (Geary Institute)
Link to online version
http://www.ucd.ie/economics/research/papers/2003/WP03.19.pdf
Item record/more information http://hdl.handle.net/10197/939
Downloaded 2021-12-05T00:00:55Z
The UCD community has made this article openly available. Please share how this access
benefits you. Your story matters! (@ucd_oa)
© Some rights reserved. For more information, please see the item record link above.
CE NTRE FOR E CONOMIC RE SE ARCH
WORKING PAPE R SE RIE S
2003
Functional Literacy, E ducational Attainment and E arnings:
A Multi-Country Comparison
K evin J Denny, Colm P Harmon and
Vincent O’Sullivan, University College Dublin
WP03/19
August 2003
DE PARTME NT OF E CONOMICS
UNIVE RSITY COLLE GE DUBLIN
BE LFIE LD DUBLIN 4
Functional literacy, educational attainment
and earnings: a multi-country comparison*
KEVIN J. DENNY
COLM P. HARMON
VINCENT O’ SULLIVAN
University College Dublin
th
Version 2.2, 25
July 2003
Abstract:
In this paper a rich and innovative dataset, the International Adult
Literacy Survey, is used to examine the impact of functional literacy on
earnings. We show that the estimated return to formal education is
sensitive to the inclusion of literacy: excluding it biases the return to
education in many countries by significant amounts. Literacy itself has a
well-determined effect on earnings in all countries though with
considerable variation in the size of the effect. The benefits of literacy
do not only arise from increasing low levels of literacy: increases at
already high levels generate substantial increases in earnings in some
countries. In general we find little interaction between schooling and
literacy though for a few countries they appear to complement each
other.
* Material from the International Adult Literacy Survey is used with permission of Statistics Canada who bears
no responsibility for the calculations contained herein or for any interpretation made by the authors. This work
forms part of the Policy Evaluation programme of the Institute for the Study of Social Change at University
College Dublin. Denny is also affiliated with the IFS, London and Harmon with CEPR & IZA. Harmon
acknowledges the award of the Nuffield Foundation New Career Development Fellowship at University College
London. Corresponding author: Dr. Kevin Denny, Department of Economics & Institute for the Study of Social
Change, University College Dublin, Belfield, Dublin 4, Ireland. Phone (+353 1) 716 4613. Fax (+3531) 716
1108, kevin.denny@ucd.ie .
1
1. Introduction
This paper examines the effect of functional literacy on individual earnings using the IALS.
Since there may be numerous determinants of earnings we estimate multivariate models allowing in
particular for the impact of formal education. There is an enormous econometric literature estimating
the impact of education. This is usually referred to as the “return to schooling” since it was shown by
Mincer in his classic 1974 study that the estimated marginal effect of a year’s education was
equivalent, under certain conditions to the internal rate of return treating education as an investment
by the individual1.
Over the 1990’s there has been a sustained and sizeable increase in the return earned by
college graduates relative to the less educated contributing to increased inequality. One explanation
for this is that increased globalisation of markets has put downward pressure on the wages of low
skill workers because of competition with low wage economies. An alternative view is that changes
in the workplace, particularly information technology, have put a premium on the skills required to
make best use of that technology this is referred to as “capital, skilled-labour complementarity”.
Distinguishing between these two explanations is important but difficult and is beyond the remit of
this paper2. Either way it is clearly important to know what the returns to skills are if policy makers
are to make the correct decisions about the provision of training.
There is a much smaller body of research estimating the effects of characteristics such as
innate abilities like “intelligence”. This is partly due to the lack of suitable data, especially outside
the United States. Furthermore in much of the economic literature on skills, “high skill” is actually
defined according to the highest education level completed and not by direct measure.
1
The most important assumption is that there is no direct cost of tuition.
2
This has, to some extent, changed in the last ten years or so partly reflecting the controversy
caused by Herrnstein & Murray’s monograph The Bell Curve (1994) but also by changes in labour
markets particularly, but not exclusively, in the US. The Bell Curve argued for the importance of
increasing importance of innate skills in the labour market. Those that possessed these skills formed
the “cognitive elite” and policies which aimed to increase the non-innate skills of the outside the elite
were judged to be unproductive. These results are hotly disputed by leading scholars, for example
after a careful examination of US data, Ashenfelter & Rouse (2000) find no evidence that that the
return to schooling varies with the measured ability of individuals.
In all this of this discussion we have used the terms like ability, skills, cognitive skills rather
loosely and interchangeably but one needs to be quite clear that there are different underlying
concepts. One can usefully distinguish between innate abilities such as intelligence and acquired
abilities such as literacy. The former are those with which one is born and are presumably stable over
time. On the other hand skills are something that can be acquired through education and training but
it seems likely that they are correlated with some inherited abilities: “smart” people are likely to find
it easier to acquire additional skills or may better appreciate the benefits of it. Measures of innate
ability should therefore be assessed early in the life of the individuals before they are “corrupted” by
educational and other interventions. Furthermore this is explicitly not what the tests were designed to
do. Measures of skills such as those used in this paper clearly will reflect, to some extent, the innate
ability of an individual but should not be interpreted as a general measure of intelligence. Indeed the
term “cognitive skill”, used in Pryor & Schaeffer (1999) study based on NALS, reflects the
somewhat ambiguous position of these tests.
2
Krussell, Ohanian, Rios-Rull & Violante (2000) suggest that capital-skill complementarity can explain much of the
variation although Denny, Harmon & Lydon (2002) find evidence to the contrary.
3
In the next section we provide an overview of the relevant literature on the effect of education
and skills on earnings. Section 3 describes the IALS data and how we use it to model earnings.
Section 4 contains the empirical analysis while section 5 concludes.
2. Modelling the effect of education and functional literacy on earnings
Studies on the determinants of earnings are typically based on standard Mincer (1974) loglinear earnings equations show that returns to education are around 6% to 8% per school year for men
3
. A useful extension to the core model is to consider the role of the individual’s ability on the
schooling decision whilst preserving the basic idea of the Mincer model of schooling being an
investment.
Griliches (1977) introduces ability explicitly into the derivation of the log-linear earnings
function. In the basic model the internal rate of return (IRR) of schooling is partly determined by
foregone income (less any subsidy such as parental contributions) and any educational costs.
Introducing ability differences has two effects on this basic calculus. The more able individuals may
be able to ‘convert’ schooling into human capital more efficiently than the less able, and this raises
the IRR for the more able. One might think of this as inherent ability and education being
complementary factors in producing human capital so that, for a given increment to schooling, a
larger endowment of ability generates more human capital. On the other hand, the more able may
have higher opportunity costs since they will typically have greater earnings potential. If ability to
progress in school is positively correlated with the ability to earn, this reduces the IRR.
Moreover, empirically least squares estimation requires that the explanatory variables are
uncorrelated with the unobserved disturbance term in the equation.
4
If an individual’s ‘ability’ or
motivation affects earnings but is omitted from the earnings equation the estimated return to
schooling will be biased.
The extent of the bias will be determined by the correlation between
education and ability. The approaches adopted to deal with this issue typically include explicit
measures for ability to proxy for unobserved ability (Blackburn & Neumark, 1993). IQ and other
such tests are an example of such proxies (Griliches (1977), Griliches & Mason (1972)). The results
of these studies have largely found favour with the notion of upward bias in least squares results.
Griliches (1977), using NLSY data which includes test scores from IQ tests conducted in high school
in addition to a survey-specific test initiated at the first interview, finds significant reduction in the
estimated return to schooling once ability measures are included. More recent studies by McKinley
Blackburn and David Neumark (1993, 1995) suggest a similar finding. Again using NLSY data they
find the OLS estimates to be some 30-40% higher when ability measures are excluded.
Finally
Murnane, Willett & Levy (1995) use mathematics ability as a regressor and find evidence of upward
bias of between 31-52%.
Boissiere, Knight & Sabot (1985) find that the return to education drops by two-thirds, once
cognitive skills are taken into account. In addition they find that this result holds albeit on a smaller
scale for manual and non-manual workers, suggesting that proficiency in literacy is essential for
productivity in all job markets. Cawley, Conneely, Heckman & Vytlacil (1996) find that a measure of
general intelligence calculated using the technique of Principal Components does not significantly
reduce the variance associated with wage regressions and the return to cognitive achievement is low
relative to the return to education, experience and family background. They also find that the choice
of occupation is determined by factors other than cognitive skills.
3
See Harmon, Oosterbeek & Walker (2003) and Heckman, Lochner & Todd(2003) for recent overviews of the literature
on the return to schooling .
5
The decision of whether to use years of schooling or highest level of education
completed is partly a matter of interpretation and to some extent a matter of taste. In the conventional
human capital model additional years in education add extra human capital so years of schooling are
the appropriate variable. With either a signalling or credentialist model it makes sense to include
measures of the highest level of education completed. In practice it is often difficult to distinguish
between such approaches empirically and the present paper makes no attempt to do so, and
frequently the implied rates of return from the two approaches give similar results (where, for
example the return to a primary degree is often worth about three of four years worth of education)4.
Moreover using years of schooling facilitates comparisons with the extensive international literature
on the subject.
The basic Mincer model to be estimated is therefore
ln y = β 0 + β S S + β X X + e
(1)
The dependent variable is the natural logarithm of earnings, S is years of schooling and X is a
set of control variables including a quadratic in age to allow for the concavity of wages with respect
to experience. The estimated β’s can then be interpreted as, approximately, the proportionate effect
on earnings of a one-unit change in the corresponding variable. Our second specification augments
(1) by adding the IALS measure of ability, denoted A, as discussed in the next section:
ln y = β 0 + β S ' S + β A A + β X X + e
(2)
The return to schooling when controlling for ability is denoted βS’ . In some cases we
estimate a variant of (2) where the ability measure is normalised within each country to have a mean
of zero and a standard deviation of one with a corresponding parameter of βAN . This will not change
the estimated value of βS’
4
.
See Krueger & Lindahl (1999) for evidence in favour of the linear-in-schooling model. Denny & Harmon (2001) use the
IALS data and find evidence of "sheepskin" effects.
6
3. The IALS Data Set & Ability Measures
The International Adult Literacy Survey (IALS) was administered by twelve governments in
association with the European Union, the OECD and UNESCO in a series of waves between 1994
and 1996. A further wave in 1998 added eight more countries5. The purpose of the survey was to
measure the literacy level of the adult population and to provide a common mechanism that would
allow comparison of literacy proficiency across countries rather than a mere count of the number of
‘illiterate’ people in the population. However it is clear from the study design that the definition of
literacy was not intended to be focused solely on comprehension, rather is was aimed at
encompassing a broad range of skills used in the context of working, schooling and home duties
which are much more cognitive in nature than the term ‘literacy’ at first suggests (OECD 1997). In
other work it has been shown how performance on the test can be predicted by educational
attainment (Denny, Harmon, McMahon & Redmond (1999)).
The survey consists for most countries of a sample of 2000 to 3000 from the adult civilian
population aged between 16 and 65. The language of interview is each country’s respective national
language. Sample design was the responsibility of each country. The IALS is structured around three
stages.
Firstly, each individual was required to complete a background questionnaire, which
provided information on age, sex, education, labour market experiences and literacy related activities.
An individual was deemed to be an IALS respondent if they partially or fully completed the
background questionnaire.
Stage 2 involved the completion of 6 simple assignments; if the
respondent answered incorrectly on more than two of these tasks the interview was terminated. This
5
The countries involved initially were Australia, Canada, Belgium, Germany, Ireland, Netherlands, New Zealand,
Sweden, Switzerland, Great Britain, Northern Ireland, United States and Poland. The two main language groups in
Switzerland and Canada were collected separately. Belgium refers to Flanders only. The final wave added Chile, the
Czech Republic, Denmark, Finland, Hungary, Italy, Norway, Slovenia as well as the Italian speaking Swiss.
7
was in order to avoid re-interviewing those individuals of whom it is known that their literacy level is
already very low (known as Level 1). Lastly a main booklet of tasks was given to each respondent
which resulted in a score, which measured their literacy level.
All assignments required the
respondent to use materials from everyday life. For example, instructions from medicine bottles, the
completion of order forms and reading a newspaper listed amongst the tasks that were required in
order to complete the test questionnaire.
The literacy level is measured on three scales: prose, document and quantitative. Prose
literacy is the knowledge required to understand and use information from texts, such as newspapers,
pamphlets and magazines. Document literacy is the knowledge and skill needed to use information
from specific formats, for example from maps, timetables and payroll forms. Quantitative literacy is
defined as the ability to use mathematical operations, such as in calculating a tip or compound
interest. In order to provide an actual measure of literacy each individual was given a score for each
task, which varied depending on the difficulty of the assignment. Scores for each scale ranges from
0-500, which is subsequently subdivided into five levels. Level 1 has a score range from 0-225 and
would indicate very low levels where, for example, instructions for a medicine prescription would
not be understood. The interval 226-275 defines Level 2 where individuals are limited to handling
material that is not too complex and clearly defined. Level 3 ranges from 276-325 and is considered
the minimum desirable threshold for most countries while Level 4 (326-375) and Level 5 (376-500)
show increasingly higher skills which integrate several sources of information or solve complex
problems.
As an example a task that involves reading the dosage on a medicine label falls into level 1 of
the prose scale, whereas a level 4 task in the prose scale may require the respondent to answer
questions from a set text. A level 1 task in the Quantitative scale may require the interviewee to add
up the total number of goods ordered from an order form, while a level 4 task on the same scale may
8
ask the respondent to calculate the total return from a compound interest table on a certain amount
(OECD 1997). To be classified at a particular level the respondent had to answer uniformly at that
level. The criterion used for consistent performance was 80 per cent. The lower the score the
respondent received at each scale, the lower the level and hence the lower one’s measure of literacy
at that scale.
In constructing the scores each country was instructed to re-score 20 per cent of tests with a
95 per cent degree of accuracy to guarantee precision of results. In addition, to ensure good quality
inter-country scores a different country re-scored 10 per cent of another country’s scores. The IALS
were also very conscious of non-response bias. Interviewers were advised to return to households
that did not give a response as many times as possible and the sample was carefully weighted to
known population variables. The survey makes uses of "plausible value" sampling methodology
which provides five measures of each of the three variables (prose, document and quantitative
literacy) based on the fact that individuals will answer different parts of a given question. Given that
each of the five is equally plausible we use the simple average to construct measures of prose,
document and quantitative ability.
IALS provides us with a unique opportunity to analyse this issue in a comparative context.
However estimation of earnings functions for the IALS data is complicated as the income data for
most countries is only observed to fall in a certain interval on a continuous scale. IALS wage data is
constructed on the basis of assigning individuals to the appropriate quintile of the wage distribution,
providing a 5-category banded income variable. Stewart (1983) shows that better estimates are
available by exploiting a distributional assumption for the continuous but unobserved variable with a
maximum likelihood estimator than ad hoc procedures such as using the mid-points of the wage
bands.
In this framework the unobserved continuous wage data is mapped into the discrete observed
9
income bands. Some observations are left-censored - we know that the unobserved income is less
than or equal to an observed censoring value. Similarly some observations are right censored - the
unobserved income is less than or equal to an observed censoring value. The estimator is a natural
generalisation of estimation of the censored normal which is in turn a generalisation of the well
known Tobit estimator. For the 1998 wave of countries the data includes continuous measure of
(annual) wages as well as the banded data. For consistency we use the banded data. If we use the
continuous data for these countries the results are very similar.
Note that our earnings data specifies which of five bands the individuals annual labour
market earnings are. The top category is unbounded. Using data on hours worked per year (which
varies across individuals and is measured continuously) we can estimate a model for hourly earnings,
where effectively the bands will vary across individuals. Estimation proceeds under the assumption
that hourly earnings are log-normally distributed which is generally found to be a reasonable
assumption (with the possible exception of the upper tail which might be better characterised by a
Pareto distribution). We also calculate robust asymptotic standard errors using the well-known
method associated inter alia with Huber & White (see Gould & Sribney (1999) for details of
estimation and computation).
Aside from the complications due to the estimation of a model with a banded dependent
variable, the model is relatively standard. Our estimates are based on a standard linear earnings
function where the earnings is expressed as a function of age and its square, dummy (binary)
variables reflecting immigrant status, whether an individual lives in an urban or rural area and the sex
of the individual and the variables of interest years of schooling completed and a single measure of
functional literacy.
Our measure of functional literacy is simply the average over the three types of literacy:
prose, document and quantitative. An alternative would be to use principal components e.g. to extract
10
the first component from all 15 plausible values and use this as a measure. This gives virtually
identical results since the weights within the component are almost the same; the correlation between
the average over all 15 and the first component is about .986. Given the richness of the data one
obvious question is whether one can fully exploit the information and measure the separate effects of
the three types of functional literacy. Including the three separately never gives sensible conclusions:
we think this because of the high correlation between them so we just the average over all three. This
raises a deeper question of whether there are three dimensions to functional literacy and if there are
whether the tests distinguish between them. This issue is not pursued further here but we note
Reder’s (1998) analysis of the US’ National Adult Literacy Survey, one of the precursors to IALS,
casts doubt on whether those tests identified distinct types of functional literacy.
Table 1 shows standard descriptive statistics for the full sample and the sample used in the
econometric analysis7.
They show that for the most parts the sample used in the analysis is very
similar to the overall sample. The proportion in rural areas (defined in IALS as living in a community
with a population of 20,000 or less) is about 1% less in the full sample. The biggest difference is the
proportion of young (16-25 years) people in the working sample which is significantly lower than the
overall sample largely because many of them are continuing in education.
6
This is consistent with the results for Canadian IALS in Green & Riddell (2003).
We have not analysed the Australian data since the public use sample provided to us excludes it. Poland is excluded, as
we were unable to discover the values defining the wage bands. The Italian speaking sample for Switzerland is also
omitted since the population is numerically small.
7
11
4. The results
The estimates of the return to education from these simple earnings equations are presented in
Figure 1. Here we summarise the earnings returns from education from our basic specification both
excluding and including our ability measure. The data presented in the table is sorted in ascending
order of the differences in the return from including ability. The returns for a number of countries are
not well known but in many respects are consistent with more general cross-country findings
including those in the meta study in Denny, Harmon & Lydon (2002). For example less developed or
transition economies tend to have higher returns to education and this is borne out in Figure 1.
However what appears to be the most interesting aspect of this figure is the quite dramatic
drop in the return when ability is included in some countries and that in particular the countries at the
bottom end of the table where the impact of including ability is low or insignificant are all from nonEnglish speaking countries. That the return to schooling falls with the inclusion of ability reflects the
fact those with higher education will in general have higher literacy so omitting literacy exaggerates
or biases upwards the estimated impact of schooling.
The alternative issue is to focus on the return to ability in these countries.
To make the
coefficient of this comparable across countries we also normalise the ability score to have mean zero
and a standard deviation of one. Figure 2 presents the earnings return to this normalised measure of
ability. Again the return to ability is highest in English speaking countries (Ireland, Great Britain,
USA) in general
However while these returns are significant it should be noted that the shift in scores required
to move an individual one standard deviation is significant. Moreover the presentation of these
results is sensitive to the distribution of the test scores. If all countries had normally distributed test
scores of similar dispersion Figure 2 could be interpreted along all parts of the score distribution.
12
However if there is variation in the dispersion of the scores or skewness in the distribution this may
not hold.
Figure 3 shows the earnings return from moving from the 25th to the 50th percentile (or
median) of the score distribution in terms of the equivalent number of years of schooling – in other
words we illustrate how many years of schooling is required to equate the earnings return of a move
from the 25th to 50th percentile in the test score distribution.
This might be thought of as the key
issue from a policy maker’s perspective. The return to a year of schooling is clear and well defined
measure. If a policymaker is trying to counteract low schooling by later interventions to raise basic
skills they need to know that that a basic skills programme is going to reap dividends particularly
given the variance in returns to these skills across the distribution of test scores as shown in Figures
2. Figure 3 is not reassuring on these matters given the wide variation in the number of years of
schooling required to equate to a standard deviation shift in scores. It again would appear, with the
exception of the Netherlands, that the English speaking countries stand to gain the most from raising
the skills level from low to median levels. For example individuals in the USA who can make this
move receive an earnings return equivalent to almost three schooling years. For Nordic countries on
the other hand, a significant shift in scores of this type only equates to under a year of schooling8.
There is little evidence that countries with low returns to education have low returns to ability
or vice versa. In fact the direct effects seem largely independent of each other across countries. The
correlation between the two sets of coefficients is not statistically significant.
8
The same picture but for the move from the median to the 75th percentile (or the 75th to the 90th ) shows, as might be
expected, slightly smaller values for the number of years of schooling equivalent measure. However the changes are
insignificant and the rank order of countries is largely unchanged.
There is little evidence that countries with low returns to education have low returns to ability or vice versa. In fact the
direct effects seem largely independent of each other across countries. The correlation between the two sets of
coefficients is not statistically significant.
13
The preceding discussion is largely based on an assumption of linearity in scores in terms of
the impact on earnings. Extending the policy implication outlined above might query whether policy
should be directed at individuals with very low levels of functional literacy or be directed across the
distribution of literacy scores. To address this we allow for non-linearity in the impact of literacy on
earnings by using dummy variables for each quintile of the IALS score distribution.
The results,
summarised in Figure 4, are most interesting. The first coordinate in each chart gives the return in
moving the first to the second quintile, i.e. from very low to low levels of functional literacy. The
return from this move can be as low as 5% for an individual, and is more typically in the region of
10-15%. However the United States in particular has an enormous gain from escaping this bottom
level of literacy of the order of 30%.
In many countries movements up the distribution of literacy scores continues to reap
dividends – movement from the second to third quintile in Great Britain is more rewarding than this
earlier transition (i.e. the marginal return to literacy is increasing at this portion of the distribution).
In general however the gains appear modest through these middle quintiles. This changes when we
examine the gains from very high levels of functional literacy – the transition into the top quintile
seems to generate quite substantial gains. In some countries (Ireland, GB, Netherlands for example)
this return is of the order of 15% and compared with the bottom quintile, individuals at the top levels
earn almost 50% more.
Throughout this specification formal schooling is also controlled for, thus these gains are
even more surprising and certainly indicate that basic skills, as measured by this IALS score, is an
important target for individual gains and perhaps therefore policy attention. More specifically
helping individuals to make transitions into the highest levels of functional literacy can make as
much difference to their earnings as moving from the lowest to next level. This may be counterintuitive because skills such as are measured in the IALS are typically labelled “basic skills” so there
14
may be a presumption that while some minimum level of these skills pays rich dividends that there is
little or no premium to increasing the skills of someone who is already highly skilled. Clearly this is
not true for some countries.
So far we have assumed that the returns to education and functional literacy are independent
of each other. This is clearly a strong – and testable- assumption which has major implications. For
example the authors of The Bell Curve argue that the returns to education were lower for those with
lower innate ability, those outside the cognitive élite. Consequently they conclude that “…school is
not a promising place to try to raise intelligence or to reduce intellectual differences“ (p 414).
However Ashenfelter & Rouse (2000) using the US’ National Longitudinal Survey of Youth find that
the returns to earnings does not vary with ability as measured by the Armed Forces Qualification
test9. Cawley, Heckman, Lochner & Vytlacil (2000) find that increases in the college premium in the
US over the 1990’s are associated with those of higher ability. As discussed earlier, the tests in IALS
are not pure measures of innate ability or “intelligence” and will partly reflect the age, education and
labour market experiences of individuals so this paper cannot address these important issues directly.
We examine this by re-estimating our basic specification but including an interaction term for ability
and years of education; the results are shown in table 4. This amount to modifying equation (2):
ln y = β 0 + β S ' S + β A A + β SA. .S . A + β X X + e
(3)
It follows that the marginal return to schooling depends on literacy and vice versa, for
example:
∂ ln y
= β S ' + β SA. . A
∂S
9
(4)
Of course the publication of The Bell Curve has generated an enormous scholarly and public debate with many of its
conclusions heavily criticised on either theoretical or empirical grounds. The use of AFQT as a general measure of
cognitive ability is disputed in Fischer, Hout, Jankowski, Lucas & Swidler (1996), chapter 3.
15
If βSA>0 the return to schooling increases with the ability level in which case ability and
schooling are complements, each enhances the marginal (proportionate) effect of the other. If βSA<0
they are substitutes. By assuming log-linearity we are imposing a form of complementarity since
even with no interaction term the marginal effect on the level (as opposed to the log) of each variable
on the level depends on the other. For example rearranging (4) and setting βSA=0 implies:
∂y
= β S ' . y ( A, S , X ...)
∂S
(5)
There appears to be a general presumption that complementarity will prevail, or failing that,
that the interaction should be zero. However it is not difficult to think of circumstances in which
these two variables would be substitutes. Consider an employer who wants an employee to possess a
set of skills some of which are imparted by formal schooling. If there is some upper-bound to the
overall level of skill required then as the employee gets closer to this limit from one skill source then
the marginal return to the other is likely to diminish. For example, an individual who has the required
skills and formal education to be a bus-driver is unlikely to increase his productivity and hence his
earnings in that occupation from gaining a university degree. The substitutability may only hold
“locally”, for a given job, since after some point additional human capital or skills may translate into
a better job.
The estimates of the parameters of interest in equation (4) are presented in Table 3. Since the
magnitude of the interaction parameter is very small (as schooling multiplied by ability is
numerically large) we multiply the estimated βSA (and its standard error) by 100. The first thing to
note from Table 3 is that most of the interactions are not statistically significant. However one needs
to take into account the joint significance of sets of variables to draw interesting inferences, in all
cases the three are jointly significant. There are in fact four distinct groups of countries in these
results.
16
For Norway, the US and the two Canada samples: the interaction term is significant and the
main effects are not jointly significant. We define this as “strong complementarity”: the marginal
benefit of either is proportional to the other. In other words only the educated benefit from having
higher ability and vice versa. The second “group” consists of Ireland, There the interaction term is
negative and significant and the main effects are jointly significant so the marginal effect of each
diminishes with the other. This is the substitutability argument discussed earlier10.
For a large group of countries (Chile, Czech Republic, Finland, Italy, Netherlands, New
Zealand and Slovenia) the main effects are jointly significant but the interaction is not, in which case
one can infer that the marginal returns are independent.
For the remaining countries (Belgium, Denmark, Germany, Great Britain, Hungary, Northern
Ireland, Sweden and the two Swiss samples), while the interaction is not statistically significant, the
two direct effects of literacy and schooling are not jointly significant either. This does not imply that
they are not matter, it may well be the case that the data is not informative enough to allow us to
estimate all of the parameters precisely or it maybe there is an interaction but it takes some other
form. Given that we expect the two variables to be correlated is not surprising. This issue is
addressed by Cawley et al (2000) in the context of using AFQT. Essentially one is unlikely to see
many individuals who have high education and low ability or vice versa. Therefore the interaction
between the two may not be identified: assuming linearity as we have done “solves” the problem but
may be a very strong assumption. This is not as likely to arise with the IALS skill measures precisely
because they are taken later in life than traditional cognitive measures so it is quite
10
In principle this could imply negative marginal returns to education or ability but this only happens for values of the
variables that are off the scale.
17
possible to observe individuals with say low education but high skills. For example In Table 4 we
show the proportion of a given ability quartile achieving a given level of education for Ireland and
Hungary. One can see that in general there are relatively few observations in the “corners”, Ireland is
to extent unusual because one observes a significant number in the highest ability quartile who have
quite low education.
5. Conclusions
This paper estimates the effects of basic (or cognitive) skills on individual earnings for a large
18
number of countries. Wages are not the only mechanism through which functional literacy may
affect an individuals labour market chances11. We have taken employment as given but it seems
plausible that the probability of an individual being employed may also depend on their cognitive
skills (Rivera-batiz (1992) and Raudenbush & Kasim (1998)). However, in general the study of
the effect of the skills measured in these tests on economic behaviour, while growing rapidly, is
in its relative infancy by comparison with our understanding of factors such as education, trade
unions or training. Therefore we still have a lot to learn about how best to model the relationship.
Our empirical results may be summarized as follow: including measures of ability or
functional literacy lowers the return to formal education in general and substantially in some
countries. Turning to the estimated effect of functional literacy itself, the effects vary substantially
across countries but in general are quite large.
In all but two countries (Slovenia and the Czech Republic) a one standard deviation increase
in literacy increases wages by more than a year of schooling does. For some countries increasing an
individual’s ability from the 25th percentile to the median has the equivalent effect on wages as
around two years of education. These countries are mostly English speaking. Allowing for ability to
have a non-linear effect on earnings presents a mixed picture. In most countries, being in the second
quintile of the distribution generates a sizeable wage premium over the first. Movements within
intermediate quintiles of ability do not always generate higher wages while for some countries there
is a substantial wage premium to being at the top of the ability distribution.
Allowing for interactions between ability and education there is also a mixed picture. For
many countries it is difficult to identify separately the direct and indirect effects and for many others
the marginal return of each is independent of the other. For a small number of countries there is
11
Literacy is also likely to influence other outcomes, Denny (2003) shows using the same data that functional literacy is
19
evidence of complementarity and in one case there is evidence that education and skills are to some
extent substitutes.
Typically researchers have assumed in the absence of other data that the patterns one finds in
a small number of countries, such as the US and Great Britain, hold more widely. One of the
strengths of the IALS is that by having internationally comparable data for a large number of
countries, it is possible to find these patterns. The very richness of the data implies that that there are
many other angles that could have been explored. Nonetheless the results show that functional
literacy has a vitally important but variable role to play in the determination of individual earnings.
an important influence on an individual’s social capital, measured as participation in voluntary and community activities.
20
References
Ashenfelter, O., & Rouse, C. (2000). Schooling, intelligence, and income in America. In K. Arrow, S.
Bowles, & S. Durlauf (Eds.), Meritocracy and economic inequality (pp. 89-117). Princeton, N.J: Princeton
University Press.
Blackburn, M.L., & Neumark, D. (1993). Omitted-ability bias and the increase in the return to schooling.
Journal of Labour Economics, 11(3), 521-544.
Blackburn, M.L., & Neumark, D. (1995). Are OLS estimates of the return to schooling biased downward?
Another look. Review of Economics & Statistics, 77(2), 217-30.
Bossiere, M., Knight, J.B., & Sabot, R.H. (1985). Earnings, schooling, ability and cognitive skills. American
Economic Review, 75(5), 1016-1030.
Cawley, J., Conneely, K., Heckman, J., & Vytlacil, E. (1996). Measuring the effects of cognitive ability.
National Bureau of Economic Research Working Paper Series no.5645.
Cawley, J., Heckman, J. Lochner, L., & Vytlacil, E. (2000). Understanding the role of cognitive ability in
accounting for the recent rise in the economic return to education. In K. Arrow, S. Bowles, & S. Durlauf
(Eds.), Meritocracy and economic inequality (pp. 231-265). Princeton, N.J: Princeton University Press.
Denny, K., & Harmon, C. (2001). Testing for sheepskin effects in earnings equations: Evidence for 5
countries. Applied Economics Letters, 8(9), 635-37.
Denny, K., Harmon C., & Lydon, R. (2002). Cross-country evidence on the returns to education: Patterns
and Explanations. CEPR Discussion Paper 3199.
Denny, K., Harmon, C., McMahon, D., & Redmond, S. (1999). Literacy and education in Ireland. Economic
and Social Review, 30(3), 215-226.
Denny, K. (2003). The effects of human capital on social capital, Institute for Study of Social Change,
University College Dublin, working paper 2003/09
Fischer, C.L., Hout, M., Jankowski, M.S., Lucas, S.R., Swidler, A., & Voss, K. (1996). Inequality by design:
Cracking the Bell Curve myth. Princeton, New Jersey: Princeton University Press.
Gould, W., & Sribney, W. (1999). Maximum likelihood estimation with Stata®. College Station, Texas: Stata
Press.
Green, D.A., & Riddell, W.C. (2003). Literacy and earnings: an investigation of the interaction of cognitive
and unobserved skills in earnings generation. Labour Economics, 10(2), 165-184.
21
Griliches, Z. (1977). Estimating the returns to schooling: Some econometric problems. Econometrica, 45(1),
1-22.
Griliches, Z., & Mason, W.M. (1972). Education, income, and ability. Journal of Political Economy, 80(3),
S74-S103.
Harmon, C., Oosterbeek, H., & Walker, I. (2003). The return to education: Microeconomics. Journal of
Economic Surveys, 17(2), 115-155.
Heckman J., Lochner L. & Todd P (2003) Fifty years of Mincer earnings regressions, NBER Working paper
9732
Herrnstein, R., & Murray, C. (1994). The Bell Curve. New York: Free Press.
Krueger, A.B., & Lindahl, M. (1999). Education for growth in Sweden and the world. Swedish Economic
Policy Review, 6(2), 289-339.
Krussel, P., Ohanian, L.E., Rios-Rull, J-V., & Violante, G.L. (2000). Capital skill complementarity and
inequality: A macroeconomic analysis. Econometrica, 68(5), 1029-1058.
Mincer, J. (1974). Schooling, experience and earnings. New York: Columbia University Press.
Murnane, R.J., Willett, J.B., & Levy, F. (1995) The Growing Importance of Cognitive
Skills in Wage Determination. Review of Economics and Statistics, 77(2), 251-266.
O.E.C.D. (1997). Literacy Skills for the Knowledge Society. Paris: OECD.
Pryor, F.L., & Schaeffer, D. (1999). Who’s not working and why: Employment, cognitive skills, wages and
the changing U.S. labor market. Cambridge: Cambridge University Press.
Raudenbush, S.W., & Kasim, R.M. (1998). Cognitive skills and economic inequality: From the National
Adult Literacy Survey. Harvard Educational Review, 68(1), 33-79.
Reder, S. (1998). Dimensionality and constructive validity of the NALS assessment. In M.C. Smith (Ed.),
Literacy for the twenty-first century: Research, policy, practices and the National Adult Literacy Survey (pp.
37-57). Westport, CT: Praegar.
Rivera-batiz, F.L. (1992). Quantitative literacy and the likelihood of employment among young adults in the
United States. Journal of Human Resources, 27(2), 313-328.
Stewart, M.B. (1983). On least squares estimation when the dependent variable is grouped. Review of
Economic Studies, 50(4), 737-753.
22
Table 1: Descriptive statistics:
Whole Sample
23
Working Sample
Mean:
Std. Dev:
Mean:
Std. Dev:
Age
Schooling
Ability
37.87
12.63
283.64
11.83
3.35
52.64
38.78
12.62
285.71
11.05
3.18
51.11
Age Intervals:
15 years of age
16 – 25
26 – 35
36 – 45
46 – 55
56 – 65
0.07
18.39
27.18
26.94
19.37
7.87
0.03
0.39
0.44
0.44
0.40
0.27
0
13.61
28.78
28.92
21.25
7.45
0.00
0.34
0.45
0.45
0.41
0.26
66 or older
0.18
0.04
0
0.00
Woman:
Rural:
Immigrant:
47.94
33.51
6.25
0.50
0.47
0.24
46.87
32.7
6.24
0.50
0.47
0.24
Father's Education:
No schooling/isced 0:
isced 1:
isced 2:
isced 3:
isced 5:
isced 6/7:
7.32
23.11
31.51
23.44
5.25
9.38
0.26
0.42
0.46
0.42
0.22
0.29
6.88
22.97
33.18
23.54
5.02
8.41
0.25
0.42
0.47
0.42
0.22
0.28
Sample size
32,002
24,978
Table 2: estimated returns:
Return to
Return
Standard
Standard
Ability
to
Ability
Error (Normalised)
Error
Return
Return
Standard
Standard
to
to
Schooling Error Schooling Error
βS
β S`
β S - β S`
βA
β AS
Belgium
0.0832
0.0079
0.0697
0.0089
0.0135
0.1823
0.0555
0.0845
0.0089
Canada
(English)
Canada
(French)
Chile
0.0881
0.0082
0.0715
0.0094
0.0166
0.1777
0.0511
0.0894
0.0094
0.1013
0.0121
0.0806
0.0151
0.0207
0.2230
0.1012
0.1068
0.0151
0.0932
0.0071
0.0679
0.0089
0.0252
0.2717
0.0598
0.1526
0.0089
Czech
0.1172
0.0109
0.1011
0.0117
0.0161
0.2355
0.0700
0.1033
0.0117
Denmark
0.0598
0.0050
0.0489
0.0055
0.0109
0.1923
0.0420
0.0712
0.0055
Finland
0.0459
0.0050
0.0351
0.0055
0.0108
0.1988
0.0443
0.0783
0.0055
Germany
0.0536
0.0082
0.0445
0.0085
0.0091
0.2256
0.0611
0.0931
0.0085
Great Britain
0.1020
0.0076
0.0747
0.0080
0.0273
0.3422
0.0410
0.1796
0.0080
Hungary
0.0911
0.0112
0.0774
0.0124
0.0136
0.1934
0.0791
0.0819
0.0124
Ireland
0.0812
0.0083
0.0551
0.0088
0.0262
0.3275
0.0470
0.1734
0.0088
Italy
0.0524
0.0058
0.0416
0.0066
0.0109
0.1519
0.0459
0.0772
0.0066
Netherlands
0.0480
0.0051
0.0353
0.0054
0.0127
0.3123
0.0507
0.1196
0.0054
New
Zealand
Northern
Ireland
Norway
0.0589
0.0072
0.0391
0.0075
0.0198
0.3063
0.0412
0.1450
0.0075
0.1114
0.0079
0.0881
0.0082
0.0233
0.2992
0.0405
0.1625
0.0082
0.0641
0.0056
0.0535
0.0064
0.0105
0.1550
0.0461
0.0624
0.0064
Slovenia
0.1520
0.0111
0.1361
0.0126
0.0159
0.1448
0.0578
0.0811
0.0126
Sweden
0.0361
0.0062
0.0280
0.0065
0.0081
0.1746
0.0462
0.0783
0.0065
Switzerland
(French)
Switzerland
(German)
USA
0.0563
0.0086
0.0473
0.0092
0.0090
0.1912
0.0728
0.0802
0.0092
0.0584
0.0109
0.0414
0.0116
0.0170
0.2707
0.0716
0.1297
0.0116
0.0834
0.0064
0.0545
0.0071
0.0290
0.3023
0.0373
0.1951
0.0071
Controls used in interval regressions: female, rural, immigrant, & dummies for father's education using ISCED levels.
(1) Not controlling for ability
(2) Controlling for ability
(3) Ability in units of 100
24
Table 3 Schooling, ability and an interaction term:
Schooling:
Standard
Ability:
Standard
Interaction:
Standard
βS`
Error:
βA
Error:
β AS
Error:
Belgium
4
0.0811
0.0017
0.2337
0.0355
-0.3943
1.3792
Canada (English)
1
0.0007
0.0012
-0.1124
0.0213
2.4124
1.1418
Canada (French)
1
-0.0627
0.0035
-0.3431
0.0612
5.0442
2.0278
Chile
3
0.0726
0.0006
0.2906
0.0121
-0.2052
1.0047
Czech
3
0.2248
0.0054
0.7596
0.0995
-4.1141
2.4071
Denmark
4
-0.0038
0.0011
-0.0231
0.0201
1.7807
1.1190
Finland
3
0.0765
0.0012
0.3600
0.0196
-1.3679
1.1269
Germany
4
0.0009
0.0027
0.0607
0.0420
1.4633
1.7362
Great Britain
4
0.0653
0.0022
0.3044
0.0353
0.3134
1.5217
Hungary
4
0.0466
0.0032
0.0520
0.0707
1.1682
2.0987
Ireland
2
0.1294
0.0015
0.5935
0.0199
-2.6142
1.3019
Italy
3
0.0804
0.0007
0.3027
0.0116
-1.4738
0.9521
Netherlands
3
0.0914
0.0012
0.5369
0.0212
-1.8800
1.1419
New Zealand
3
-0.0298
0.0017
0.0316
0.0282
2.2787
1.3513
Northern Ireland
4
0.0985
0.0022
0.3421
0.0388
-0.3431
1.5397
Norway
1
-0.0165
0.0013
-0.1143
0.0208
2.3060
1.1706
Slovenia
3
0.0721
0.0018
-0.1291
0.0342
2.5857
1.6590
Sweden
4
0.0503
0.0012
0.2535
0.0168
-0.7229
1.1098
Switzerland (French)
4
0.1197
0.0024
0.5055
0.0492
-2.5257
1.6819
Switzerland (German) 4
-0.0337
0.0021
-0.0435
0.0400
2.7269
1.6235
USA
0.0038
0.0005
0.0549
0.0121
1.8555
0.7781
1
Note: The numbers in the last two columns have been multiplied by 100. 1 beside country name implies
interaction significant, main effects not. 2 implies main effects and interaction significant. 3 implies main
effect significant but interaction not. 4 implies main effects not significant and interaction not.
25
Table 4: Cross tabulation of schooling by ability for 2 countries:
Ireland:
Schooling:
Hungary:
Q1
Q2
Q3
Q4
Q1
Q2
Q3
Q4
59.53
32.68
22.96
9.73
31.28
14.23
7.32
2.44
31.52
49.81
49.81
38.52
51.85
58.54
58.13
31.3
8.17
16.34
24.9
42.41
13.99
21.95
24.8
44.31
0.78
1.17
2.33
9.34
2.88
5.28
9.76
21.95
5 to 9 years
10 to 12 years
14 to 17 years
16+ years
26
27
itz
e
rla
S
nd wed
( F en
re
n
G ch )
er
m
an
Fi y
nl
an
d
Ita
l
D
en y
m
a
N rk
or
N
w
Be
lg ethe ay
iu
r
l
m
a
(F nds
la
Sw
nd
itz
er
s
er
la Hun )
nd
ga
(G ry
Ca erm
n a an
da )
(E
n
Sl g)
o
N
ew ven
Ze ia
al
an
d
Ca Cz
n a ec
N
h
or
d
th a (
Fr
er
n
e
Ire )
la
nd
Ch
ile
Ire
la
nd
G
U
re
S
at
B A
rit
ai
n
Sw
Rate of Return (%)
FIGURE 1
RATES OF RETURN TO SCHOOLING
0.16
Schooling - NO ABILITY
Schooling - ABILITY
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
28
D
or
w
a
en y
m
ar
k
Fi
nl
an
d
H
un
ga
ry
Ita
ly
Sw
Sw
ed
itz
e
er Slo n
la
nd ven
(F ia
re
C
Be an nch
lg ada )
iu
m (En
(F
g
la )
nd
e
G rs)
e
Ca rma
n
na
da y
(F
re
)
Sw
Cz
ec
itz Ne
th
er
e h
la
nd rlan
(G d s
e
N rm
ew an
Ze )
al
an
d
Ch
ile
N
or
th Irel
er
n an d
I
G rela
re
at nd
B
rit
ai
n
U
SA
N
Return to Ability (Normalised Measure) - %
FIGURE 2
RATES OF RETURN TO ABILITY (normalised)
0.2
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
Note: This figure shows βAN from Table 2
FIGURE 3 YEARS OF SCHOOLING EQUIVALENT TO MOVE FROM 25TH TO 50TH PERCENTILE
3.5
3
Years of Schooling
2.5
2
1.5
1
0.5
Sl
ov
en
Be
ia
H
lg
u
iu
m n ga
r
(F
la y
nd
er
N s)
or
w
a
D
en y
Ca
m
n a ark
da
Ca (E
n a ng)
da
Sw
(F
re
itz
)
er
Cz
la
nd
e
( F ch
re
nc
G h)
er
m
an
N
y
or
th
er Ita
n
l
Ire y
la
nd
Fi
Sw
nl
an
itz
d
er
la
C
nd
hi
(G
le
er
m
an
)
Sw
ed
en
G Irela
re
nd
at
B
rit
ai
n
U
N
S
et
he A
N rlan
ew
d
Ze s
al
an
d
0
Note: The graph above solves for S in the equation below, where S is the number of years schooling of
an individual and Q50 and Q25 are the 50th and 25th percentile of the ability measure respectively. βS’
and βA are the estimated returns for each country taken from Table 2 as described in equation (2).
βS' S = βA(Q50 −Q25)
29
Figure 4
Return to Quintiles of Ability (first quintile omitted)
Canada (English)
Belgium (Flanders)
.4
.3
.2
.1
.6
Effect on Wages
.5
0
.5
.4
.3
.2
.1
0
20
40
60
80
Percentile of Ability
100
Chile
40
60
80
Percentile of Ability
.4
.3
.2
.1
100
20
Czech Republic
.4
.3
.2
.1
0
.5
.4
.3
.2
.1
0
40
60
80
Percentile of Ability
100
40
60
80
Percentile of Ability
100
100
40
60
80
Percentile of Ability
100
.6
Effect on Wages
Effect on Wages
.5
40
60
80
Percentile of Ability
Denmark
.6
20
.5
0
20
.6
Effect on Wages
Canada (French)
.6
Effect on Wages
Effect on Wages
.6
.5
.4
.3
.2
.1
0
20
40
60
80
Percentile of Ability
100
20
Finland
Effect on Wages
.6
.5
.4
.3
.2
.1
0
20
The graphs show the estimated coefficient on a dummy variable representing the quintile of the ability distribution of an individual. The vertical
access represents the proportionate effect on wages i.e. 0.1= 10%
30
Figure 4 (cont.)
Return to Quintiles of Ability (first quintile omitted)
Germany
Great Britain
.4
.3
.2
.1
.5
.4
.3
.2
.1
0
20
40
60
80
Percentile of Ability
Ireland
60
80
Percentile of Ability
.3
.2
.1
20
100
.4
.3
.2
.1
0
100
New Zealand
.6
.5
.4
.3
.2
.1
0
40
60
80
Percentile of Ability
100
100
Netherlands
.5
.4
.3
.2
.1
0
40
60
80
Percentile of Ability
40
60
80
Percentile of Ability
.6
Effect on Wages
.5
20
.4
Italy
Effect on Wages
Effect on Wages
40
.6
20
.5
0
20
100
.6
Effect on Wages
.6
Effect on Wages
.5
0
31
Hungary
.6
Effect on Wages
Effect on Wages
.6
.5
.4
.3
.2
.1
0
20
40
60
80
Percentile of Ability
100
20
40
60
80
Percentile of Ability
100
Figure 4 (cont.)
Return to Quintiles of Ability (first quintile omitted)
Norway
.5
.5
.5
.4
.3
.2
.1
.4
.3
.2
.1
0
20
40
60
80
Percentile of Ability
.3
.2
.1
40
60
80
Percentile of Ability
20
100
Switzerland (French)
.6
.5
.5
.5
.3
.2
.1
0
Effect on Wages
.6
.4
.4
.3
.2
.1
0
20
40
60
80
Percentile of Ability
100
40
60
80
Percentile of Ability
100
USA
.6
.5
.4
.3
.2
.1
0
20
40
60
80
Percentile of Ability
100
Switzerland (German)
.6
Effect on Wages
Effect on Wages
.4
0
20
100
Sweden
Effect on Wages
Effect on Wages
.6
0
32
Slovenia
.6
Effect on Wages
Effect on Wages
Northern Ireland
.6
.4
.3
.2
.1
0
20
40
60
80
Percentile of Ability
100
20
40
60
80
Percentile of Ability
100