Human capital externalities and return to education in Turkey Preliminary Version Ozan Bakis
∗
Haluk Levent
†
Sezgin Polat
‡
May 25, 2009
Abstract This paper studies local human capital externalities and return to education in Turkey. The data comes from 2006 Household Labor Survey. Ordinary Least Squares estimates yield 1.8 % of human capital externalities while Instrumental Variables estimates are higher: 3.4 %. We discuss further implications of social return to education for the segmented labor market by gender and job contract-term divide. Our results imply that in a segmented labor market, the social return to education is more pronounced for less educated segment of the local labor market. Keywords: human capital externalities, returns to schooling JEL Classi cation Numbers: I20; J24; D62.
Galatasaray University Galatasaray University ‡ Galatasaray University, This research project was sponsored by Galatasaray University Project Fund. We would like to thank Galatasaray University for giving us the nancial support for this research. ∗ †
1
1 Introduction Human capital and research and development (R&D) are the main determinants of growth according to the new endogenous growth theories. In Romer (1986, 1990), Lucas (1988) and Aghion and Howitt (1992) the mechanism by which human capital is the source of growth diers slightly from one work to other. Spillovers, or external eects of human capital underly at least if one does not take into account semantic concerns the same thing: factor payments are higher than their marginal productivity of workers. Hence, the amount of human capital and knowledge is vital for a society. In real life, one of the biggest part of human capital, R&D or knowledge is schooling. Further, numerous empirical works suggest that formal schooling
1
is an important determinant of productivity levels . Dierences in human capital are seen the main reason of the income disparities as pointed by Mankiw et al. (1992) and Krueger and Lindahl (2001). Return to education can be decomposed into two parts: individual returns analyzed by Mincerian approach and social returns due to externalities developed by the endogenous growth theory. Even if there is not yet consensus on how large are social returns almost all economists accept that they should exist intuitively. After all, what would be the major contribution of the endogenous growth theory based on externalities otherwise? In this paper we are interested in social returns of education in Turkey. Our objective is policy oriented: given the high ratio of young population in Turkey what should be the priorities of the future Turkish governments in education sector? Is it desirable that all young people have college degree? Given various problems of being a developing country, especially the lack of nancial resources, it is primordial to know where to invest an extra Turkish Lira (TL) and what is to be expected from this investment. This is why the existence and amplitude of social returns need to be investigated. We use the 2006 Household Labor Survey for our analysis. We exploit the distinction of term structure of job contract to segregate the labor market in Turkey. This divide enables us to analyze additionally the impact of social return to education on structural composition of the local labor markets. We also discuss the implications for each gender separately. Ordinary Least Squares (OLS) estimates show positive and important local human capital externalities due to average level of schooling in the region. In order to estimate social returns to schooling correctly we need to
1
For instance, Romer (1989), Mankiw et al. (1992), Benhabib and Spiegel (1994), Barro (1999), Krueger and Lindahl (2001) and Aghion et al. (2004) nd that schooling is positively correlated with GDP per worker. 2
estimate private returns as well. According to OLS estimates an extra year of schooling is associated with 6.8 % increase in private returns; similarly 1% point increase in regional college share raises average wages by 1.8% points. The major problem about measuring social returns to education in empirical studies, is the endogeneity problem. Average wage in regions may be high because of at least three reasons; only one if them is the one we search: human capital externalities. The second one is unobserved city characteristics correlated with local human capital stock that make workers more productive. Consider two regions, A et B. If an unobserved factor like amenity but any omitted variable correlated with unobserved individual heterogeneity can be taken is higher in A; and skilled workers prefer amenity, then, skilled workers will ceteris paribus choose region A to work/live in. As skilled workers will be more productive, average wages in region A will be higher even without any externality eect. In order to get unbiased estimates of human capital externalities we need to control for such regional characteristics correlated with local human capital of the region.
The third one is un-
observed individual characteristics correlated with individual and regional human capital level that makes wages higher. Take the example of Silicon Valley. The industrial structure of Silicon Valley increases wages of talented workers and therefore the ratio of workers with high ability will be higher in the Silicon Valley. If we do not include a control variable for individual ability our estimates of social return externalities will be biased. Therefore, higher average wages in a region does not re ect necessarily human capital externalities.
There can be unobserved individual/regional
variables that make wages higher in that region. In order to control for these unobserved eects, we will use Instrumental Variables (IV) method.
We
need at least two instruments; one for individual and the other for regional human capital levels. These instruments must be correlated with individual and regional human capital but uncorrelated with unobserved factors that aect wages. We use household size of the family as an instrument for individual human capital.
In Turkey, family size seems to be inversely correlated with
education level probably because of some socioeconomic reasons. But there is no reason for ability to be correlated with family size. Ratio of working women in total employment is our second instrumental variable: this ratio is correlated with level of human capital in the region but uncorrelated with unobserved regional characteristics that may increase wages. IV estimates yield similar results: 1% point increase in education raises wages by 11.6 % points and 1 % point increase in college share is associated with 3.4 % point increase in average wages.
3
Early literature about human capital externalities is concentrated on cross-country regressions (e.g., Romer (1989), Benhabib and Spiegel (1994) Barro (1999), Krueger and Lindahl (2001) and Aghion et al. (2004)). These papers are especially interested in whether it is stock or accumulation of human capital that creates externalities.
Rauch (1993) claims that high
wages can be associated with a high level of economic development which is, in return, associated with higher levels of capital stock. So, it would be very di cult to identify the eects of human capital externalities in cross-country data. This is why we have an emerging body of literature that studies a single
2
country . Studying only one country implies that externalities in question are local human capital externalities.
This is already in parallel with the
pioneer paper of Lucas (1988) who stated explicitly the role of cities as center of spillover eects when he discusses the importance of interaction between dierent agents. The rst paper on local human capital externalities is Rauch (1993). Using data from Standard Metropolitan Statistical Areas (SMSAs) in the USA, and generalized least squares method he nds that there are local external eects of average schooling, while average level of experience is not signi cant. Rauch does not deal with endogeneity problem of human capital, as he uses a random eects model to take into account that the eect of unobserved city characteristics. One major problem of this method is that it assumes random eect not to be correlated with other explanatory variables. Acemoglu and Angrist (2000) nd similar results to the ones obtained by Rauch (1993) when they use OLS estimates.
But, endogeneity prob-
lem of human capital leads the authors to make the same regression using instrumental variables and then they nd no signi cant of human capital externalities. They use state compulsory attendance laws and child labor laws (in individuals' state of birth when they were 14) as instruments of average human capital of the states, because these laws are correlated with future human capital averages and are exogenous to future adult wages.
Their
results are based on a sample of white men aged 40-49 from the 1960-80 Censuses. An interesting point of the paper is the use of quarter of birth as an instrument for individual schooling. Moretti (2004b) uses instrumental variables in order to solve endogeneity problem. Used instruments are age structure of cities and presence of landgrant colleges in cities.
The basic idea is that younger cohorts are better
educated than older ones and presence of land-grant colleges established
2 See especially Rauch (1993), Acemoglu and Angrist (2000), Moretti (2004) and Ciccone and Peri (2006)
4
more than 100 years ago is an instrument for human capital.
He founds
that a one percentage point increase in college share, after controlling for private return, raises average wages by .6%-1.2%. In addition to papers mentioned above there are two related papers that estimate the private return of education for Turkey. The rst one is by Tansel (2004). The paper studies the private return to education in Turkey from two sources: 1987 Household Expenditure Survey (HES) and 1989 Household Labor Force Survey (HLFS). Using HES, it estimates the private return of education for each education level
3
for men and women separately. The
method is joint maximum likelihood estimation of equation and the wage equation.
For men, Tansel(2004) nds a rate of return between 1.90 and
16.20 for wage earners (all ages included). Secondly, it uses a two-step estimation method of Heckman and the HFLS data. Results for wage earners are very close to the 1987 HES ones, while for self-employed people, the range is between 6.14 and 14.70. The second one is Guner and Duygan (2005) who use 2002 Household Income and Consumption Expenditures Surveys (HICES). They nd that one extra year of education increases earnings by 12.57% on average in a standard Mincerian approach. Their dependent variable is logarithm of the annual wage earnings
4
and independent variables are: years of schooling,
years of experience, years of experience squared. They use a sample of all males between the ages of 20-54 in the 2002 HICES for their regressions. The papers by Tansel (2004) and Guner and Duygan (2005) are important in their contribution of private returns to education. However, social returns to education are not studied for the Turkish case. Our paper can be considered as a rst attempt studying human capital externalities for Turkey.
2 Model The underlying model is adapted from Rauch (1993) and Moretti (2004a). The model that will be developed will enable us to test for externality of human capital in cities. In order to have positive externalities, we need that in some developed regions wages are higher than less developed ones. One can ask why wages are not similar across the country? The answer is based on the assumption that thanks to migration toward regions with high wages, we will have higher residential and commercial rents in these regions. As a
3
Identi ed levels are primary, middle, high, vocational high schools and university. The use of total wage is problematic; in fact one needs a good measure of marketed education return; i.e. hourly wages. 4
5
result, the utility level will become almost equal across all regions. A worker will be indierent between high wages/high rents and low wages/low rents. Consider a country with land
ti
N
regions.
A region
and a local public good (or an amenity)
i = 1, . . . , N
do not pay for, with
i has a xed amount of ai for which households
denoting regions. There are many rms
and households who can migrate at zero cost between regions; and many households with dierent levels of human capital. A household human capital (or equally, e cient labor)
wi ,
at wage per e ciency unit,
in region
hj
j has a level of
which she supplies inelastically
i.
There is a single and nationally traded consumption good
y
produced
by capital, labor and land under perfect competition. Consumption good's price is normalized to unity, inputs:
p = 1.
labor, land and capital.
Returns to scale are constant in private Representative rm in region
i
has the
following production function:
yi = A(hi )F (ki , hi , ti ) A(hi )
is the externality eect that depends on the aggregate level of human
capital at the region
i.
Individual rm does not have control on it.
important point is that the rental price of capital
r
The
is common to all regions
while the prices of land and labor are region speci c.
The reason is that
land and labor are traded locally, while capital not. Land's rental price is denoted
zi
in region
i.
Preferences are all identical and homothetic across households. Households get utility from land, amenity and consumption good. Representative household living in region
i
has the following utility function:
Ui = u(ai , yi , ti ) It is standard to derive cost function for rms and indirect utility function for households.
Under constant returns to scale and perfect competition,
unit cost is given by (we neglect
r
as it is common to all regions),
p = 1 = Ci (wi , zi , A) and indirect utility per e ciency unit of labor
Vi = V0 = v(wi , zi , ai ) The spatial equilibrium is obtained when households and rms are indifferent between regions. Common nation-wide utility of one e ciency unit
6
of labor is denoted
V0 .
Then, equilibrium is there where all rms have unit
marginal costs and all households have
V0
for one e ciency unit of labor.
Consider two regions, A and B such that amenity in region A is greater than amenity in B, i.e.
aA > aB .
We will assume that skilled workers value
amenity more than unskilled ones. Then, ceteris paribus a skilled worker will accept a lower wage in region A in comparison with region B. If there is not spillover eects due to local human capital wages of skilled workers would be lower in region A. The only way we can expect higher wages in region A for skilled workers are externalities associated with local human capital. Since workers are mobile between regions, in equilibrium, in every region there will be workers from each human capital level. If spillover eects increase nominal wages of skilled workers, say in region A, than rental prices should also be high in this region in order to ensure that workers are indifferent between any region.
3 Identi cation problem If there are human capital externalities wages in regions with higher levels of human capital will be higher. But, as we have already mentioned, wage dierentials may also stem from some unobserved heterogeneity between regions. We can regroup these unobserved factors in two groups: individual unobserved factors or region-speci c characteristics. In order to identify human capital externalities we need some instruments that are not correlated with unobserved factors that raise wages but correlated with individual and regional human capital levels. The model that we will use to estimate social returns is
log wij = β0 + Xij β1 + Hj β2 + Zj β3 + uij
(1)
X: holds for all individual attributes including age, tenure, having social security ( formal/informal) and individual human capital,the education level, H: holds for regional human capital which is proxied by college share in the total employment, Z: is the regional indicators that can be correlated with regional human capital H. The error term can seen as a linear combination of three factors: an unobserved individual factor
θ
(such as ability) that is possibly correlated with
individual human capital; an unobservable regional factor possibly correlated with level of skilled labor in that region; and nally an independently and identically distributed shock,
.
uij = µj + θi + ij 7
(2)
Endogeneity problem can arise from both unobserved individual and regional characteristics.
First, let us see how unobserved individual factors
can give rise to the endogeneity problem. It is possible that more talented workers are more productive in a certain region A say because of the technological structure that complements ability and thus earn higher nominal wages there. As a result, we expect the ratio of skilled workers to be higher in region A, even if all of the observed worker characteristics are the same. In terms used in equations (1) and (2) we would say
Cov(θ, H) > 0.
Therefore,
the positive correlation between high wages and high level of human capital in region A, is not due to externality; but only to an unobserved factor: high returns to ability in region A. In order to control for this bias we need an instrument that presents exogenous variation. This instrument should be (i) correlated with individual education (human capital); (ii) uncorrelated with unobserved factor (ability). Our preferred
5
instrument for individual education levels is the family size.
We can test the rst part of the above condition; we see that family size is correlated with individual education level. But, we can not test the second part. We assume that family size is not correlated with individual's ability. Secondly, unobserved regional heterogeneity may bias our estimates. If, for some cultural, geographical or any other reason unrelated with average human capital level of the region that make workers less/more productive in that region, dummy variables will control for some of unobserved heterogeneity. The regional development index can partially capture these eects. But, evidently, the index can not account for all omitted variables that correlate wages with average human capital in the region. In the above notation we would write
Cov(µ, H) 6= 0.
As Acemoglu and Angrist (2000, p.2) put:
Rauch's estimates are driven by dierences in average schooling across cities. But higher incomes might cause more schooling instead of vice versa. Cities with greater average schooling may also have higher wages for a variety of other reasons. This highlights the fact that a major challenge in estimating the eects of education on income is identi cation. To solve this endogeneity problem due to unobservable regional factors, we will use ratio of working women ratio as an instrument of local human capital stock. As before, it is possible to test whether our proposed instrument is correlated with regional human capital.
5
In order our instrument
We would like to use panel data, as in Moretti (2004b), to follow the same individual over time and see how changes in levels of local human capital change individual's wage. Given lack of panel data this is not possible for Turkey. 8
to be valid, it should not be correlated with unobserved factors that aect wages but correlated with local human capital level. We nd that the ratio of working women in total employment is correlated with college share with 75.1%.
4 Regression analysis In order to isolate the market price of education we use only hourly wages as our dependent variable in all regressions. As underlying theoretical suggests, the dependent variable is nominal wages in our model.
The parameter of
interest is coe cient of the average human capital in regions. We tried to show the impact of social return to education on two dierent type of workers namely for regular employees who work on regular basis and casual (temporary) employees who work rather on daily basis. Regular workers have by de nition of Turkstat have longer term job-contracts and earn monthly income.
The causal employees rather work on daily basis,
have shorter and job-speci c term-contracts. This segmentation of work is important for the Turkish labor market since taking into consideration the regional disparities, in some regions, we have a very dierent contract-term composition. Our OLS estimations shed some light to this fact when private return to education is also taken for granted. We begin the analysis with this structural divide in the labor market. The results indicate a sharp decline for private return to education for casual workers. In the OLS estimations, private return to education is 6.8% for both total and men in the regular workers' group and whereas private return decline to 1.8% for total 1% for men in the casual workers' group. Our social return to education variable, the college share of workers in total employment have a positive sign and signi cant for all OLS estimates. Social return to education is 1.8% and 1.6% for total and men respectively in the regular workers' group. Compared to regular work, the social return to education is higher for the total having a 1.9% eect and declines for men to 0.8% for men in the casual work group. We can otherwise say that 1% point of increase in the college share raises average wages by 1.6 % points for men and by 1.8 % points for the total in the regular work group. The same reasoning holds for the coee cents of the causal work group. The development index which take account the amenities of the region is negative and signicative for all models except for men in the casual work group. The negative coe cent of the development index imply that everything holds constant, for a given wage, workers prefer to accept
9
10
0.068*** 0.038*** -0.000*** 0.362*** 0.037*** -0.001*** 0.018*** -0.113*** -1.060*** 65857 0.495
Education Age Age2 Social Security Tenure Tenure2 College Share Reg. Dev. Index Cons
No. Obs. R2 adjusted
Source: TURKSTAT
* p
