This article was downloaded by: [University of Sydney Library]

On: 15 August 2015, At: 01:40

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: 5 Howick
Place, London, SW1P 1WG

Applied Economics
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/raec20

Linearity of the return to education and self

selection
Marta Sanmartín
Published online: 09 Nov 2010.

To cite this article: Marta Sanmartín (2001) Linearity of the return to education and self selection, Applied Economics,
33:1, 133-142

To link to this article: http://dx.doi.org/10.1080/00036840121772

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”)
contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy, completeness, or suitability for any
purpose of the Content. Any opinions and views expressed in this publication are the opinions and views
of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content
should not be relied upon and should be independently verified with primary sources of information. Taylor
and Francis shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses,
damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connection
with, in relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
 Applied Economics, 2001, 33, 133± 142

Linearity of the return to education and self

selection
M A R T A S A N M A R T IÂ N
Departamento de Fundamentos del AnaÂlisis EconoÂmico. Avda Lehendakari Agirre 83,
CP 48015 Bilbao, Spain, E-mail: [email protected].
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

This paper investigates the earnings-education pro® le in Spain. First no functional

form is imposed on this relation but instead dummy variables are used for each year
of education. Second, diŒerent alternatives are analysed until the best parametric
form is found. Moreover problems related to the self selection of individuals and to
the segmentation of the labour market based on the educational level are considered.
The main results indicate that the eŒect of education on earnings is not signi® cant
until individuals ® nish secondary education and from there onwards the relationship
can be considered linear.

I. INTRODUCTION Variations on this speci® cation can arise if we assume

that the return to education does not remain constant as
Following human capital theory, we can say that education we increase the educational level. Accepting, for instance, a
gives a higher production capacity to individuals; hence, in quadratic functional form would give rise to convex or
a competitive framework, individual earnings must rise concave earnings-education pro® les, which would imply
with educational level. In this sense, many studies have that the return to an additional year of education is
tried to estimate the returns to education, this is to say, increasing or decreasing, respectively. A more recent re-
the eŒect of an additional year of education on earnings. search line has tried to capture potential deviations from
The model used by most studies is the one developed by linearity by eliminating the restrictive linear parameteriza-
Mincer (1974), who proposes a linear relation between the tion used by Mincer. In this sense, in order to allow for
logarithm of earnings and the number of years of educa- greater ¯ exibility in earnings-education pro® les, some
tion. Heckman and Polachek (1974) oŒer empirical evi- studies have replaced the traditional linear functional
dence in favour of the log-linear functional form against form by a dummy variable for each year of education.
alternative models. Their model has been used with data on Following this method, Hungerford and Solon (1987)
diŒerent countries and various periods of time, getting ® nd evidence of the diploma eŒect for graduates. Park
returns to education up to 16% for each additional year (1994) also captures a deviation from linearity at this par-
(Ashenfelter and Krueger, 1994). For Spain, Alba-Ramõ Â rez ticular moment. In spite of this, both studies accept a pro-
and San Segundo (1995) estimate a return to education of ® le close to linear. Gri n and Cox Edwards (1993) get
8.1% , and Corugedo et al. (1992) get an estimation of 6.9% convex earnings± education pro® les for Brazil.
for secondary education in Spain. 1 Furthermore, when estimating the returns to education
Nevertheless, this functional form can be used under using the Mincer earnings equation, education has some-
some particular assumptions, as it considers that the rate times been considered as a predetermined variable. In this
of return is the same for all educational levels. Its theor- way, problems derived from diŒerences in the learning ca-
etical justi® cation is based on the assumption of homo- pacity of individuals, from the ® nancial capacity to pay the
geneity of the human capital acquired by individuals. cost of learning, or from any other factor which aŒects not

1
These authors assume a mixed model of human capital and signalling. The literature of signalling asserts that education per se does not
increase individual’s productivity, but helps to send a signal to the employer about his/her innate capacity. Willis (1986) is an excellent
survey of human capital literature. Psacharopoulos (1985) also oŒers information about the empirical results for returns to education.
Applied Economics ISSN 0003± 6846 print/ISSN 1466± 4283 online # 2001 Taylor & Francis Ltd 133
http://www.tandf.co.uk/journals
 134 M. SanMartõ Â n
only the returns to education, but also the decision to edu- empirical analysis. Finally, section IV is devoted to the
cate, are not taken into account.2 Thus, there has been a most important conclusions.
great deal of attention paid to the demand for education.
As human capital theory suggests, individuals choose the
educational level which maximizes the discounted present
value of their lifetime earnings. As Griliches (1977) notes, I I . TH E D A T A
`. . . schooling is the result, at least in part, of optimizing
behaviour by individuals and their families. This behavior The data used in this study are taken from two diŒerent
is based on some anticipated earnings function’ . sources. The ® rst one is the survey named `Encuesta de
Socioeconomic variables, such as the environment where Estructura, Conciencia y Biografõ Â a de Clase (ECBC)’.
individuals grow up, the ® nancial situation of their This survey was carried out in 1991 by the National
families, and individuals’ innate capacity and perspectives Bureau of Statistics (INE), the local government of
as regards to future earnings are determinant for the deci- Madrid and the Spanish Women’ s Institute. It contains
sion concerning educational level. In order to test the rele- information on 6632 individuals.
vance of these variables, Rodrõ Â guez (1992), Beneito et al. This survey is structured in two sections: in the ® rst one,
(1995) estimate equations of demand for education for each interviewed person answers questions related to his/
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

Spanish households derived from family utility functions. her job, and family situation. The second part of the survey
The existence of some characteristics which aŒect the indi- is about personal opinions of the individuals interviewed.
vidual educational level would mean that education is not All data used in this paper are taken from the ® rst part of
randomly distributed. The probability of reaching a par- the survey.
The sample used in the paper presents a drawback: the
ticular educational level may be higher for some individ-
lack of equiprobability. Some individuals are more likely to
uals, and this must be taken into account when estimating
be interviewed than others. For instance, individuals living
earnings.3
in Madrid are overrepresented, as 1600 out of 6632 inter-
The aim of this work is to study the earnings± education
views are carried out there. Young, highly educated indi-
pro® le for Spain. We want to see if the log± linear relation-
viduals are also overrepresented. However, the weight to
ship between earnings and education is empirically
solve these problems is provided, so weighted data are used
accepted, or if such standard parameterization is not
to obtain a random sample.4
valid for estimating the returns to education correctly.
The ® nal sample, after eliminating some inconsistencies
The ® rst stage is not to impose any functional form on and missing values, covers 2128 salaried workers aged
the relationship between education and earnings. Instead, between 18 and 65, whose net monthly earnings are
we use dummy variables for each year of education. By known. 5
observing the results derived from the estimations, we In order to see if the result found using ECBC is robust
look at diŒerent alternatives for adjusting one speci® c func- to the use of other data bases, an additional survey is used.
tional form. This is the `Pilot Earnings Survey’ which is an annex of the
The results indicate that the eŒect of education on earn- Spanish Current Population Survey (second quarter of
ings is not signi® cant until individuals ® nish secondary 1990). The Current Population Survey covers about
education. From there onwards, the relationship can be 60 000 households each quarter. Nobody is asked about
considered linear, and each additional year of education his/her earnings. However, in the second quarter of 1990,
implies an increase in earnings of 6.7% . This result seems 2000 households randomly taken from the outgoing rota-
to be quite robust, since it does not change when we allow tion group were chosen to be asked about their earnings. In
for the earnings structure to vary with the level of educa- 643 households, there were no members in the labour force.
tion, considering at the same time that individuals self- From the remaining 1357 households, 1259 reported infor-
select into diŒerent levels of education. mation on earnings. The ® nal sample covers 1194 salaried
The study is organized as follows: in the next section, the workers aged between 18 and 65, whose gross monthly
data used in the paper is presented; Section III presents the earnings are known.

2
Therefore, the returns to education we get if we estimate the earnings equation developed by Mincer by OLS can be biased if more
capable individuals study more, and at the same time are more productive workers. In this case, part of the estimated returns would be
due to other individual characteristics.
3
Willis and Rosen (1979) develop a model of demand for education where each occupation requires speci® c characteristics. Hence,
individual capacity is not only intelligence, but also manual ability, physical strength, etc. The result they get is that individuals self-select
to educational levels which maximize their earnings.
4
For more information, see CarabanÄa et al. (1992).
5
It is assumed that those individuals who report their earnings are a random sample. Were they a diŒerent group of individuals, there
might be a selectivity problem.
 Return to education and self selection 135
Table 1. Means and standard errors the use of the Pilot Survey annex to the Spanish Current
ECBC Pilot survey
Population Survey.
The ® rst stage consists of estimating an earnings equa-
Hourly earnings* 636 (405)** 715 (419) tion derived from the standard model developed by
Age 36.5 (11.36) 38.2 (12.41) Mincer, where it is assumed that the returns to education
Education (years) 9.85 (4.10) 7.98 (4.57)
Primary 47.24% 67.25% are constant for each additional year of education.
Secondary 34.33% 18.43% Therefore, the results shown in column (a) of Table 2 are
University 18.43% 14.32% the estimations of the following equation:
Experience 18.19 (12.52) ±
Married*** 0.61 0.64
ln Y ˆ ¬ ‡ ­ S ‡ ¸1 X ‡ ¸2 X 2 ‡ ¯Z ‡ u …1†
Sex*** 0.34 0.32
Public*** 0.22 0.27
Father’s education 4.87 (4.60) ± where ln Y represents the log of hourly net earnings; S is
Mother’s education 3.56 (3.55) ± years of education; X is the variable which measures the
Number of observations 2128 1194 postschooling human capital accumulation; and Z is the
Notes: * For the ECBC the hourly net earnings are presented. For vector of other explanatory variables.6
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

the Pilot the hourly gross earnings. ** Numbers in brackets are It can be seen (Table 2, column (a)) that the estimation
the standard errors. *** Dummy variables which take the value of of the returns to schooling is 3.5% for each additional year.
one when the individual is married, he/she is female or he/she is However, if it was decided to introduce education in the
working in the public sector.
earnings equation not linearly but in a quadratic form, it
can be seen that the returns to education diŒer as education
Table 1 shows a brief descriptive analysis of the variables increases. The result obtained when imposing this quadra-
in each of the two samples. Before starting the empirical tic form is that the pro® le of earnings-education is convex
analysis the Spanish education system that concerns the (see Table 2, column (a)). Once the individual has four years
individuals will be explained.
The primary education in Spain starts when children are
Table 2. Estimation of the earnings equation. Dependent variable:
six years old. At this moment they begin studing primary logarithm of hourly net earnings (data from ECBC)
education (EGB: Basic General Education) and it includes
eight years. With three more years individuals ® nish sec- (a) (b) (c)
ondary education (BUP) and pre-university (COU) with an
S* 0.035 (11.68)** 70.027 (73.42)
additional year. But if they want a more technical educa- S2 0.0033 (8.54)
tion when ® nishing primary, they can undertake a voca- Sˆ1 70.221 (71.54)
tional formation, FPI (two years) and FPII (® ve years). Sˆ2 70.031 (70.31)
Finally at the end of COU or FPII (and BUP in some Sˆ3 0.097 (1.15)
Sˆ4 0.173 (2.10)
cases) they start their studies at the University and it is
Sˆ5 70.066 (70.95)
possible to study for three or ® ve years (six years for Sˆ6 0.062 (0.94)
some degrees, but that cannot be distinguished in our data). Sˆ7 0.034 (0.53)
Therefore, the schooling variable goes from zero years Sˆ8 0.020 (0.41)
of education up to 18. The individual who has ® nished his/ Sˆ9 0.076 (1.21)
S ˆ 10 0.095 (1.56)
her studies with a university degree has 17 schooling years.
S ˆ 11 0.115 (1.92)
As there are very few individuals with more than university S ˆ 12 0.216 (3.70)
studies, all of them were included together in a unique S ˆ 13 0.131 (2.05)
group considering they have eighteen years of education. S ˆ 14 0.206 (2.26)
S ˆ 15 0.360 (5.97)
S ˆ 16 0.521 (3.97)
S ˆ 17 0.564 (8.85)
III. EMPIRICAL ANALYSIS S > 17 0.821 (7.80)
N 2128 2128 2128
R2 0.4118 0.4315 0.4357
3.1. Empirical analysis: search for a parametric functional
form Notes: *S is the number of years that the individual needed for the
reported education in (a) and (b) columns. In column (c) 18 dum-
The empirical analysis began by making use of the ECBC; mies are used for each year of education. **Numbers in brackets
next, it was ascertained whether the results were robust to are t-ratios

6
In Z dummies are included to control for gender, civil state, public sector, occupations (7 dummies), industry (10 dummies) and region
(17 dummies). Each of these variables is explained in detail in the data appendix
 136 M. SanMartõ Â n
of education, each additional year positively and increas- equality test was performed for each coe cient of the esti-
ingly aŒects his/her earnings. mated dummies with the corresponding coe cient in a
However, from not imposing any speci® c functional linear functional form. To that end the sample was divided
form (and hence introducing a dummy variable for each into two groups, the ® rst containing individuals with
year of education), it can be observed (Table 2, column (c)) S < 11 and the second containing individuals with S equals
that the earnings of individuals who have not achieved at 11 years or more. (It was realized that there may be selec-
least secondary education (s ˆ 12) are not signi® cantly dif- tion bias when doing this, but later on it can be seen that
ferent (at a con® dence level of 95% ) to those of individuals the returns to education do not change even considering
whose educational level is null.7 This result could be the potential selection bias problem.) The result of this test
extended to s ˆ 11 at a con® dence level of 90% .8 is that one cannot reject any null hypothesis of equality of
From s ˆ 11 onwards, it is possible to adjust a pro® le coe cients for any level of education except for s ˆ 18.
which is very similar to a linear one, considering that there This exception could be justi® ed if it can be taken into
is a positive deviation from linearity when secondary edu- account that in this group individuals with 18 years of
cation is achieved (Fig. 1 shows the pro® le resulting from education or more have been included, so that this is a
this regression). 9 fairly heterogeneous group.
One possible interpretation of the positive deviation Looking at these results, it seems reasonable to say that
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

when secondary education is completed is that those considering all the range between s ˆ 0 to s ˆ 18, neither
individuals who complete their secondary education have the linear speci® cation nor the quadratic one re¯ ects ade-
11 or 12 years of education (if they have studied the pre- quately the earnings± education pro® le. The results indicate
university course (COU)), whereas those who ® nish their that there may be two diŒerent pro® les, depending on the
years of education. In order to look at this possibility, Park
studies when they have ® nished technical± profesional edu-
(1994) and Hungerford and Solon (1987) are followed, and
cation (an equivalent to vocational education) have studied
the following speci® cation is proposed:
13 years. On the one hand, the returns to secondary educa-
tion seem to be higher than the returns to vocational edu- ln Y ˆ ¬ ‡ ­ S ‡ ® …S ¡ 11†D ‡ ¶D ‡ ¯Z ‡ u …2†
cation, but on the other hand, the group of s ˆ 13 or s ˆ 14
includes those individuals who start university but drop where ln Y represents the log of net hourly earnings; S is
out, and this might be a signal of low productivity. years of education; D is a dummy variable which takes the
value of 1 if the individual has studied at least 11 years and
Hence, this might help us to understand the positive devi-
0 if not; Z is the vector of explanatory variables; ® nally, u is
ation from linearity of those with s ˆ 11 or s ˆ 12. An
the random component:10
By doing this one is allowing for the existence of two
diŒerent pro® les, both of them linear but with diŒerent
slopes. ­^, re¯ ects the estimation of the return to education
for the ® rst 11 years and ­^ ‡ ®^ is the estimated return for
the second part. The signi® cance of ¶, ^ would show that the
level of earnings diŒers when s 11.11 The result of this
ˆ
estimation is presented in Table 3 (column (1)).
Under this speci® cation, the return to an additional year
of schooling is not signi® cantly diŒerent from zero (at 5%
of signi® cance) when the individual belongs to the less
educated group (S < 11). From this level onwards, indi-
vidual earnings increase by 6.7% for each additional year
of education. On the other hand, there seems to be no
Fig. 1. Earnings± education pro® le using dummy variables for each change in level, which means that those individuals who
year of education belong to the more educated group do not experience an

7
This result would be maintained were reference s ˆ 8, which is the level of education undertaken by the majority of individuals.
8
There is an exception when s ˆ 4. It must be noted however, that there are only 20 individuals in this educational group, which might
not be enough to get any representative result.
9
Card and Krueger (1992) observe an earnings± education pro® le very similar to this one, with the diŒerence that the returns to education
start to be signi® cantly diŒerent from zero when education is four years. Nevertheless, they do not consider this ® nding to be very
important as it aŒects only 2% of the sample. Therefore, they adjust a linear functional form. However, in our case, those individuals
whose educational level is less than 11 years represent 68% of the sample, which justi® es taking the ® nding into consideration.
10
Note that individuals with s ˆ 11 are those who have ® nished secondary education.
11
¡11®^ ‡ ¶^ gives the diŒerence in earnings level for those who belong to the group of S 11, where ¡11®^ would correspond to the
diŒerence including the slope and ¶^ is the speci® c change in the level.
 Return to education and self selection 137
earnings increase independent of the number of years of be approximated by age. Therefore one must be cautious,
education. 12 as the results might be sensitive to these diŒerences.
In Table 3 (column (2)), we present the results of the In Table 4 the results of earnings equation are presented
following speci® cation: using the linear speci® cation, the quadratic one, and that
obtained by using a dummy variable for each year of edu-
ln Y ˆ ¬ ‡ ­ S ‡ ® 1 …S ¡ 11†D ‡ ® 2 …S 2 ¡ 112†D ‡ ¶D cation. As before, for the linear case it can be seen that an
additional year of schooling increases earnings by 3.4% .
‡ ¯Z ‡ u …3† With respect to the quadratic form, a convex function
between earnings and schooling can also be seen, although
Equation 3 is estimated in order to incorporate the poss-
the degree of convexity diŒers from the one obtained using
ibility of a quadratic relation between earnings and educa-
ECBC. However column 3 shows that it is not until the
tion for the more educated group. The results reveal that
individual ® nishes eight years of study that he/she starts
the return to education is constant for each additional year
getting a positive return to education. In order to see
once the individual is in the highly educated group. This is
if there might be, as before, a change in the earnings±
so because the coe cient ®^2 is not signi® cantly diŒerent
education pro® le when s ˆ 8, the following speci® cations
from zero. Therefore, Equation 2 seems to be more ade-
are proposed:
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

quate than Equation 3 in re¯ ecting the eŒect of education

on earnings.
ln Y ˆ ¬ ‡ ­ S ‡ ® …S ¡ 8†D ‡ ¶D ‡ ¯Z ‡ u …4†
These results lead to the conclusion that linearity is a
good approach to describing the relationship between earn- ln Y ˆ ¬ ‡ ­ S ‡ ® 1 …S ¡ 8†D ‡ ® 2 …S 2 ¡ 82†D ‡ ¶D ‡ ¯Z ‡ u
ings and education, but only from a particular level of
schooling. More speci® cally, individual earnings increase …5†
at a constant rate of 6.7% for each additional year of
education once the individual has at least secondary educa- The results from these estimations are included in Table 5.
tion. Until that moment, one can consider that the return It can be seen that an additional year of schooling only
to education is null. increases earnings once the individual has ® nished primary
Given the importance of this result, it is of interest to see
if it is robust to diŒerent samples. In order to do this, an Table 4. Estimation of the earning equation. Dependent variable:
identical analysis using another Spanish data base is car- logarithm of hourly gross earnings (data from pilot survey)
ried out: the Pilot Survey of Earnings.
(a) (b) (c)

Comparative study. The main drawback that this survey S* 0:034 …9:25†** 0:033 …0:34†
presents with respect to ECBC is that whereas from S2 0:0018 …3:31†
ECBC the exact number of years studied by each individ- Sˆ5 0:045 …1:03†
Sˆ8 0:117 …2:25†
ual is known, in the Pilot Survey information about S ˆ tp 0:282 …4:59†
schooling is aggregated into nine levels. In addition to S ˆ 12 0:354 …5:81†
this, from the latter, there are fewer explanatory vari- S ˆ 15 0:471 …6:74†
ables, as there is no information neither about the region S ˆ 17 0:569 …7:74†
where individuals live, nor about experience, which will N 1194 1194 1194
R2 0:4550 0:4596 0:4612

Table 3. Dependent variable: logarithm of hourly net earnings Notes: *Since the exact number of schooling years are not known,
(data from ECBC) zero years have been given if the individual has no studies; ® ve
years if he/she answered primary education; eight years if
(a) (b) the individual said EGB (EducacioÂn General BaÂsica which is
equivalent to primary); twelve years for individuals with second-
S 0.0059 (1.17)* 0.0059 (1.17) ary education; ten years if they studied two years of Technical±
…S ¡ 11† D 0.0672 (8.35) 0.0431 (2.53) Professional education and 13 if they studied ® ve years of
…S2 ¡ 112† D ± 0.0034 (1.60) Technical± Professional education; 15 years if the individual has
D 0.0336 (1.12) 0.0553 (1.68) passed three years at the University and ® nally, 17 years if the
R2 0.4311 0.4318 individual attained a university degree. When the dummy vari-
N 2128 2128 ables are used, S ˆ tp takes the value of one if the individual has
two or ® ve years of Technical± Professional education. ** Num-
Note: * Numbers in brackets are t-ratios bers in parentheses are t-ratios.

12
A positive and signi® cant value of ¶^ would mean that those with secondary education might send signals of higher capacity (innate or
acquired) and hence higher productivity to the market. Alternatively, a positive value might also be interpreted as re¯ ecting the in¯ uence
of ability, if one thinks that ability and higher education are positively and strongly correlated.
 138 M. SanMartõ Â n
Table 5. Estimation of the earning equation. Dependent variable: Table 6. Estimation of the earnings equation. Dependent variable:
logarithm of hourly gross earnings (data from the pilot survey) logarithm of hourly net earnings (data from ECBC)

(a) (b) (a) (b)

S 0.0091 (1.03)* 0.0091 (1.04) S 0.034 (3.81)* 0.023 (2.69)

…S ¡ 8† D 0.0384 (3.51) 0.0413 (2.21) S*age 76.6 e70.6 (70.031) 0.0003 (1.65)
…S 2 ¡ 8 2 †D ± 70.0003 (70.81) age 0.0077 (2.63) 0.0017 (0.86)
D 0.0551 (1.21) 0.0534 (1.15) experience 0.0083 (2.49) ±
R2 0.4598 0.4593 experiencie2 70.00025 (74.09) ±
N 1194 1194 R2 0.4149 0.4108
N 2128 2128
Note: * Numbers in parentheses are t-ratios
Note: * Numbers in parentheses are t-ratios

education. From this moment onwards, it seems that the

ages. Given the increase in average schooling for young
relationship between earnings and education follows a
people in Spain, it might be thought that the return to
linear pro® le (note that the quadratic form cannot be
education for the young population is null, whereas it is
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

accepted). For the more educated individuals (those with

much higher than 6.7% for older workers. If all workers
more than primary education), the return to education is
are aggregated, a return to education of 6.7% as an average
3.8% . might be found. In order to see if this argument has any
As a result of these ® ndings, it can be concluded that the validity age and an interaction of age and education have
earnings± education pro® le obtained using ECBC is robust been introduced as explanatory variables; experience is
to the change of sample. This means that for Spain, it dropped (Table 6, column (1)) that the return to education
seems that the return to education is null until individuals does not vary with age. If it can be seen as an explanatory
reach a particular level of schooling, and from then variable due to its potential high correlation with age, the
onwards, the relationship between earnings and schooling result of the nonsigni® cance of the interaction between age
is linear. However, there are some diŒerences between the and education remains constant (Table 6, column (b)).
® ndings of the two samples that deserve comment. On the Therefore, this does not seem to be the explanation for
one hand, from the Pilot Survey, it is found that the mini- the two slopes in the relationship between earnings and
mum level of schooling for the individual to have positive schooling.
returns to education is primary education, whereas from
ECBC, it is necessary to have secondary education to start
getting positive returns. On the other hand, the return to DiŒerent earnings structure and self-selection
education for the group of more educated individuals is In the previous section a particular functional form for the
higher using ECBC (6.7% ) than using the Pilot Survey earnings equation was found, such that the returns to edu-
(3.8% ). The fact that from the Pilot Survey the exact num- cation must be estimated separately for two groups of indi-
ber of years of education cannot be known, as well as the viduals. However, this speci® cation assumes that the
necessity of using fewer explanatory variables (no region remaining explanatory variables have the same eŒect on
controls and experience proxied by age) might explain earnings for both groups of workers. If jobs where the
these diŒerences. Nevertheless, the fact that from both more educated workers work are substantially diŒerent
data sets it has been found that for a group of less educated from jobs where the less educated workers work, then
workers the return to education is null and that for the rest this assumption might be too restrictive, and hence estima-
the relationship between earnings and schooling is linear tions of returns to education might be biased.
seems to suggest that the ® ndings of these two diŒerent This potential problem has led us to carry out a Chow
parts are robust to a change in the sample. test of equality of coe cients in the earnings equation for
It is accepted, therefore, that the return to education is the two groups of workers. It gives a value of
null until the individual has reached a particular level of F39;2050 ˆ 1:387, which means that we reject the null of
education, and that from then onwards the relationship equality of coe cients of both subsamples at a signi® cance
between earnings and schooling is linear. Furthermore, as level of 10% . Given this result, we have estimated both
ECBC gives richer information, it has been decided to subsamples separately. From this estimation, we can see
accept the results found using that database, i.e. that the that the coe cients of the remaining explanatory variables
return to education is linear and 6.7% for those individuals are diŒerent for the two groups (see Table 7).
who have reached secondary education, and null for the In particular, the following can be seen: ® rst, the con-
rest. cavity of the earnings pro® le with respect to experience
It might be argued however, that the pro® le found in this disappears for the group of less educated workers; second,
paper is the result of aggregating individuals of diŒerent earnings do not increase if they work in a public institution;
 Return to education and self selection 139
Table 7. Estimation of the earnings equation. Dependent variable: logarithm of hourly net earnings* (data from ECBC)

High educ. Low educ. All individuals

S 0.067 (11.13)** 0.0085 (1.37) 0.035 (11.68)

sex 70.202 (77.32) 70.117 (73.48) 70.156 (77.31)
experience 0.021 (4.91) 0.006 (1.61) 0.014 (4.85)
experience2 70.00042 (74.24) 70.00013 (71.67) 70.00023 (74.09)
tenure 0.010 (4.93) 0.007 (4.16) 0.007 (5.75)
public 0.152 (4.59) 0.053 (1.05) 0.101 (3.52)
married 0.029 (0.95) 0.069 (2.06) 0.066 (2.96)
N 1278 850 2128
R2 0.4118 0.3007 0.4118

Notes: * In this table only the results for variables concerned with human capital (education, experience and tenure) are presented, a
dummy variable which takes the value of one if the individual is a female, a dummy which takes the value of one if he/she is married and
another one if the individual works in the public sector. Indicators for industry, occupation and region have also been included.
** Numbers in parentheses are t-ratios.
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

third, being female aŒects earnings negatively but the eŒect where Equations 6 and 7 are the earnings equations for the
is smaller than for the more educated group; ® nally, being more and less educated group, respectively, and Equation 8
married has a positive eŒect on the less educated group, but determines how individuals are assigned to each group.
not on the more educated one. These results provide evi- Individual decisions will depend on the earnings diŒerential
dence that the earnings structure is diŒerent for the two between the two groups and on a set of characteristics,
groups of workers. With respect to the main variable, edu- included in V , which determine whether each individual
cation, it can be seen that the result obtained concerning is assigned to the less or the more educated group. Ii¤ is a
return to education, i.e., that it is null for the less educated latent variable that assigns individuals to group 1 if Ii > 0,
group and 6.7% for the more educated one, remains the and to group 2 if Ii µ 0. Substituting Equations 6 and 7 in
same even when earnings are estimated separately for each Equation 8, the latter can be written as:
group.
Ii ˆ ³Wi ¡ ei …9†
However, it is known that if human capital theory is
followed, each individual chooses his/her educational where ei ˆ ¡® …u1i ¡ u2i† ¡ "i
level. Given some individual and family characteristics If wage determination is independent of individual
and taking into account expected future earnings, each choice, then separate estimations by OLS for each group
person decides how much to invest in education. This are unbiased, as there is no sample selection problem. 13
means that the results obtained before may be biased, as The problem arises when E …ui1 =Ii > 0† 6ˆ 0 and
the two samples are not randomly distributed among the E …ui2 =Ii µ 0† 6ˆ 0. In this case, the situation would be the
population. following:
This potential problem is taken into consideration and a
model used which considers the possibility of diŒerent E …ln Y1 =Ii > 0† ˆ E …ln Y1 =³Wi > ei† ˆ ­ 1 X1 ‡ ¼1e ¶ 1 …10†
earnings structures for the two groups of workers and
E …ln Y2 =Ii µ 0† ˆ E …ln Y2 =³Wi µ ei† ˆ ­ 2 X2 ‡ ¼2e ¶ 2 …11†
allows individuals to be included in a particular one
depending on their characteristics. This leads to the follow- where
ing switching model being considered:
¡f …³Wi† f …³Wi †
ln Yi1 ˆ ­ 1 Xi ‡ ui1 …6† ¶ i1 ˆ ¶ i2 ˆ
F …³Wi † 1 ¡ F …³Wi †
lnYi2 ˆ ­ 2 Xi ‡ ui2 …7† ¼1e and ¼2e are the covariances between the error term of
the earnings equation and the error term of the decision
Ii ˆ ® …ln Yi1 ¡ ln Yi2 † ‡ ¯Vi ‡ "i …8†
equation;14 and f …:† y F …:† are the density and distribution

13
In this case, ei would be uncorrelated with ui1 y ui2 .
14
ui1 , ui2 and ei are normally distributed with mean zero and variance and covariance matrix:
¼21 ¼12 ¼1e
¼ ˆ ¼21 ¼22 ¼2e
¼e1 ¼e2 1
 140 M. SanMartõ Â n
function of a normal with mean zero and unit variance, to observe directly the individual earnings diŒerential
respectively. derived from being in one or other educational group),
Heckman (1979) develops an estimation method to solve age (included in order to control for diŒerences in educa-
^ into the
this problem. He introduces a correction term (¶) tional level between diŒerent cohorts), some variables
earnings equation which is obtained from a maximum like- which approximate individual capacity (innate or acquired)
lihood estimation of ³; this means that the mean of the and variables re¯ ecting likelihood of schooling. The latter
error term is zero. Hence the following expressions are group includes years of education of the interview’s parents
obtained: and the social status reported by the individual (there are
three possible levels, high, medium and low social class;
ln Yi1 ˆ ­ 1 Xi1 ‡ ¼1e ¶^i1 ‡ vi1 …12†
with this variable one can try to proxy the social status
of the individual when he/she was a student, and at the
ln Yi2 ˆ ­ 2 Xi2 ‡ ¼2e ¶^i2 ‡ vi2 …13†
same time try to capture economic possibilities of each
Using Heckman two-step method we do not obtain e - individual for investing in education). The work status of
cient estimates. In order to get e ciency one must estimate the father is also included (if he is a salaried worker with an
by maximum likelihood Equations 6, 7 and 9 jointly. inde® nite or temporary contract, or self-employed with or
The variables introduced in the decision equation are all without employees). Finally indicators for the region where
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

those included in the earnings equation (as it is not possible the worker lived when he/she was 16 are included, i.e.,
Table 8. Earnings equations for both education levels and results for the selection equation. (Data from ECBC Survey)

High educ.* Low educ. Selec. function

S 0.063 (7.05) 0.008 (1.20)

sex 70.157 (73.79) 70.113 (72.71) 70.437 (74.59)
married 0.003 (0.069) 0.052 (1.319) 70.032 (70.32)
experience 0.022 (3.67) 0.0085 (1.81) 70.062 (74.48)
experience2 70.003 (72.15) 70.00014 (71.55) 70.0004 (71.40)
tenure 0.010 (3.30) 0.0081 (3.65) 70.0086 (71.36)
public 0.138 (2.87) 0.030 (0.51) 70.050 (70.37)
father’s education 0.017 (1.44)
mother’s education 0.051 (3.41)
status 0.145 (4.12)
age 0.055 (6.96)
head: temporary contract 0.299 (2.21)
head: with employees 0.119 (0.43)
head: without employees 0.219 (2.43)
head: family-subsidy 0.126 (0.32)
aragoÂn 70.287 (70.76)
baleares 0.802 (2.02)
canarias 0.520 (0.80)
cantabria 0.953 (1.56)
cleon 0.154 (0.78)
cmancha 0.509 (2.58)
catalunÄa 0.185 (1.19)
cmurcia 70.200 (70.36)
cvalencia 0.0005 (0.003)
extremadura 0.410 (1.61)
galicia 70.211 (70.65)
navarra 1.045 (2.24)
madrid 0.147 (0.82)
pvasco 0.572 (2.07)
rioja 0.251 (0.25)
¼1 0.436 (35.15)
¼2 0.371 (47.36)
»1e 70.619 (79.51)
»2e 0.257 (1.63)
N 1151 763 1914

Note: * In all equations dummies for diŒerent industries, occupations and regions where the individual lived when he/she was interviewed
are used.

15
Taking into account the de® nition of ei , in order to get a positive selection bias it is necesary for the correlation coe cient » e1 to be
negative.
 Return to education and self selection 141
when he/she ® nished secondary school. With all these vari- that for Spain the return to education is null until individ-
ables one can try to capture the intellectual capacity of each uals complete secondary education, and from then on a
individual, the background where the individual grew up linear relationship can be accepted between earnings and
and the economic facility for investing in education. education, with a return of 6.7% for each additional year
Table 8 presents the results of this estimation. With of schooling.
respect to the variables which aŒect the decision to invest
in education, it can be observed that the mother’ s educa-
tion, belonging to a high social status and some particular
situations of the head of the household positively aŒect the ACKNOWLEDGEMENTS
decision to be in the group of the highly educated. Another I thank my advisers Sara de la Rica and Arantza Ugidos
determinant is the region where the individual lived when for their helpful comments and suggestions. I also thank
he/she was 16. The estimation of the returns to education is Paula Adam for her comments. Financial support from
still null for the less educated group, and it is 6.3% (which UPV 035.321-HB230/95 and GV 035.321-0014/95 are
is not signi® cantly diŒerent from the 6.7% found before), gratefully acknowledeged.
for the group of workers who have at least secondary edu-
cation.
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

The results also indicate that there is no sample selection REFERENCES

bias for the low educated group (at 95% con® dence) Alba-Ramõ Â rez, A. and San Segundo, M. J. (1995) The returns to
whereas there is a positive sample selection bias for the education in Spain, Economics of Education Review, 14(2),
high educated group of workers.15 Hence it can be con- 155± 66.
cluded that, on average, earnings of the group of more Ashenfelter, O. and Krueger, A. (1994) Estimates of the economic
educated workers (those with secondary education or return to schooling from a new sample of twins, The
American Economic Review, 84(5), 1157± 73.
more) are higher than they would be were these workers Beneito, P., Ferri, J., MoltoÂ, M. and Uriel, E. (1995)
assigned randomly to this group. Although the point value Determinantes de la demanda de educacioÂn en EspanÄa,
of the return to education found using this model is some- Instituto Valenciano de Investigaciones EconoÂmicas,
what smaller, it is not signi® cantly diŒerent from the other Valencia, Working Paper.
one. CarabanÄa, J., GonzaÂlez, J. J., Serrano, A., Vallejos, A., Veredas,
S., RivieÁre, J. and Navarro, C. (1992) Encuesta de Estructura,
From these results it may be concluded that the results Conciencia y Biografõ  a de Clase: Informe Te cnico. Servicio de
presented before, i.e., that the return to education is null Estadõ  sticas Demogra® cas. Consejerõ  a de Economõ  a.
for workers with less than secondary education and 6.7% Comunidad de Madrid.
for workers with secondary education or more remains the Card, D. and Krueger, A. B. (1992) Does school quality matter?
same when the method of estimation is re® ned. Return to education and the characteristics of public schools
in the United States, Journal of Political Economy, 100 (1), 1±
40.
Corugedo, I., Garcia PeÂrez, E. and Martinez Pages, J. (1992)
IV. CONCLUSIONS EducacioÂn y rentas. Una aplicacioÂn al caso de la ensenÄanza
media en EspanÄa: una nota, Investigaciones EconoÂmicas, XVI
The main aim of this paper is to estimate the returns to (2), 299± 304.
Gri n, P. and Cox Edwards, A. (1993) Rates of return to educa-
education in Spain. The ® rst result obtained is that linearity tion in Brazil: do labor market conditions matter?, Economics
cannot be accepted between earnings and education once of Education Review, 12(3), 245± 56.
the individual has reached a particular educational level. Griliches, Z. (1977) Estimating the returns to schooling: some
That level is secondary education. For those who have not econometric problems, Econometrica, 45(1), 1± 22.
reached secondary education, the returns to education are Heckman, J. and Polachek, S. (1974) Empirical evidence on the
functional form of the earnings± schooling relationship,
null. Journal of the American Statistical Association, 69(346),
Another result obtained is that educational level segre- 350± 54.
gates individuals into two diŒerent groups, which indeed Heckman, J. (1979) Sample selection bias as a speci® cation error,
have diŒerent earning structures. Furthermore, the choice Econometrica, 47(1), 153± 61.
of educational level is not random but positively related to Hungerford, T. and Solon, G. (1987) Sheepskin eŒects in the
returns to education, The Review of Economics and
earnings level. In other words, the probability of being in Statistics, 69(1), 175± 77.
the group of more educated workers (secondary education Mincer, J. (1974) Schooling, Experience and Earnings. Columbia
or more) is positively correlated with unobserved factors University Press for the National Bureau of Economic
which increase workers’ productivity. However, when con- Research, New York.
sidering self-selection of individuals into educational Park, J. H. (1994) Returns to schooling: a peculiar deviation from
linearity, Working Paper no. 335, Princeton University.
groups, it is found that the empirical returns to education Psacharopoulos, G. (1985) Returns to education: a further inter-
are not signi® cantly diŒerent from those found without national update and implications, The Journal of Human
considering such self-selection. The ® nal conclusion is Resources, XX, (4), 583± 604.
 142 M. SanMartõ Â n
Rodrõ Â guez, G. C. (1992) La adquisicioÂn de capital humano: un S: Number of years needed for the educational level
modelo teoÂrico y su contrastacioÂn, Investigaciones reported by the individual.
Econo micas, XVI(2), 305± 16.
Willis, R. J. (1986) Wage Determinants: A Survey and sex: Dummy variable which takes the value of one if the
Reinterpretation of Human Capital Earnings Functions, in individual is female and zero if male.
Ashenfelter and Layard, Handbooks in Economics series, (5), married: Dummy variable which takes the value of one if
525± 602, Amsterdam, Oxford and Tokyo, North Holland, the individual is married and zero for the rest of the cases
distributed in North America by Elsevier Science, New York.
Willis, R. J. and Rosen, S. (1979) Education and self selection, (single, separated, widowed or divorced).
Journal of Political Economy, 87(5), 7± 36. tenure: Number of years that the individual has worked in
the same ® rm.
tenure2 : Squared of tenure.
A P P E N D I X : D ES C R I P T I O N O F V A R I A BL E S experience: The age of the individual is used as a proxy of
this variable.
Variables from the `Encuesta de Estructura, Conciencia y experience2 : Squared of experience.
Biografõ Â a de Clase’ (ECBC) public: Dummy variable which takes the value of one if the
ln Y : Net hourly earnings. This ® gure has been obtained individual works for the government and zero for the rest
Downloaded by [University of Sydney Library] at 01:40 15 August 2015

for all those individuals who report monthly or annual cases (public ® rm, private ® rm or others).
earnings; for those who report gross earnings, the income Finally seven dummy variables for occupations and ten
tax is discounted to get net earnings. dummy variables for industries are used.
S: Number of years needed for the educational level
reported by the individual.
sex: Dummy variable which takes the value of one if the
individual is female and zero if male. Variables used in the decision equation
married: Dummy variable which takes the value of one if
All the variables used in the earnings equations and in
the individual is married and zero for the rest of the cases
addition:
(single, separated, widowed or divorced).
father’ s education: Number of years needed for the educa-
tenure: Number of years that the individual has worked in
tion level reported by the head of the family when the indi-
the same ® rm.
experience: Age of the individual minus his/her age when vidual was 16.
he started working for the ® rst time. This variable is used mother’ s education: Number of years needed for the educa-
as one proxy for seniority. tion level reported by the mother (if she was not the head of
experience2 : Squared of experience. This variable is not the family) when the individual was 16.
signi® cative like a regressor so it is not used. status: Variable which takes the value of one when the
public: Dummy variable which takes the value of one if the individual belongs to the low social class, three if he/she
individual works for the government and zero for the rest belongs to the middle class and ® ve when he/she is in the
cases (public ® rm, private ® rm or others). high class.
Finally seven dummy variables for occupations, ten age: Age of the individual.
dummy variables for industries and 16 dummy variables work status of the father: Five categories for the work status
for regions are used. of the head of the family when the individual was 16
(inde® nite contract, temporary contract, self-employed
Variables from the `Pilot Earnings Survey’ with and without employees and receiving family-
subsidy).
ln Y : Gross hourly earnings. This ® gure has been obtained Finally, 16 dummy variables for the region where the
from the monthly earnings and from the hours that the individual lived when he/she was 16 are used.
individual uses working weekly.