You are on page 1of 6

The New Mincer-Equation

Mark den Hollander(10979603), Eefje Roelfsema (10993673),


Soufyan Boulhout (10670343) and Luuk van der Kleij (11004517)
University of Amsterdam - Faculty of Economics

June 2017

Abstract
This paper continues on the argument that the standard Mincer equation does not fit the data
quite as well as it used to. Evidence suggests that this is due to changes in contemporary data. To
characterize these changes we tested several assumptions, including semi-log and convex relations.
Both the semi-log and the convex relation seem to be significant for US data from 2011. Adjusting
the standard Mincer equation, in accordance with the theory and our test results, resulted in a
adjusted model that has a better fit on the data.

Introduction also conclude that the semi-log functional form


might be misleading in some respect for Swedish
Bjorklund and Kjellstrom (2002) introduce sev- data (Bjorklund and Kjellstrom, 2002, p. 209).
eral adjustments for the standard Mincer equa- However, Card (1999) shows stronger support for
tion that could result in a better fit for the data. semi-log on US data.
In our research we will continue on their sug-
gested adjustments and test whether we can find
a significant effect. We expect that some adjust- Data
ments need to be made to the standard Mincer
model for a better fit of the data. The data used in this paper consists of a cross
section of the U.S. Current Population Survey
from 2011. We restrict our data to only contain
Theoretical knowledge observations of hourly wage between 2 and 90.
Since we continue on the Mincer equation, which
The Mincer equation models the natural loga- focuses on the working class, all observations
rithm of earnings as a function of years of educa- with age greater than 65 are removed. Observa-
tion and years of potential labor market experi- tions with a negative work experience, which are
ence. Although The model is nearly 40 years old, doubtful, are left out because we have no further
it is still widely used. However, Lemieux (2006) information. Instead of focusing on males only,
and Bjornklund and Kjellstrom (2002) stress that as in Bjorklund and Kjellstrom (2002), we focus
the standard Mincer equation may not be a good on both males and females. The reason for this is
fit for more recent data. Lemieux (2006) argues that in recent years increasingly more woman are
that this is due to the convex function of years enrolled in higher education (Norton and Caki-
of schooling (Lemieux, 2006, p. 14). To charac- taki, 2016, p. 27). Therefore, it may be interest-
terize this convex relation, Edin and Holmlund ing to compare the human capital characteristics
(1993) extended the model with dummy vari- of both groups. Our final data set consists of
ables for years of education (Edin and Holm- 5586 observations, see table 1 for the descriptive
lund, 1993, p. 16). Bjorklund and Kjellstrom statistics.

1
Model & Results line with the theory we observe a convex relation
for years of schooling. Surprisingly, however,
Our model continues on the theoretical evidence this convexity can not be found for a doctorates
of Edin and Holmlund (1995) and Card (1999), degree. Our findings suggest that a doctorates
which assumes the use of dummy variables for degree (.771) is less valuable in terms of earnings
years of schooling and a semi-log functional form. than a masters degree (.832). An explanation
Testing the latter assumption with an LR-test might be that doctorate candidates, who work at
(table 2 ) gives us strong statistical evidence for universities, have a fixed payment. We do how-
the use of the semi-log form. Performing a Chow ever observe that a masters degree has a lower
break test resulted in a rejection of the zero- benefit from work experience (-.011).
hypothesis for equal coefficients between males As expected does work experience have a positive
and females (table 3 ). Therefore we included effect (.033) on earnings. Another observation
interaction variables for female with work expe- is that there are higher polynomial relations for
rience and female with work experience squared. work experience. However, we found no signif-
We purposely omitted the dummy FEMALE icant effect of a quartic relation, as Lemieux
because we assume that the gender wage gap is (2006) argues (Lemieux, 2006, p. 135).
merely an effect of a difference in human cap- Furthermore we observed a difference in the ef-
ital characteristics between genders. We also fect of work experience between males and fe-
observed a significant interaction effect between males. The results suggest that women receive
a masters degree and the benefit of work expe- a lower benefit from work experience (-.001) in
rience. comparison to males. However, their benefit for
The Breusch-Pagan test for heteroscedasticity work experience seems to decline less rapidly.
(table 4 ) rejects the zero-hypothesis of con- This follows our assumption that the gender
stant variation, which implies the use of White wage gap can be explained by a difference in
heteroskedastic-consistent standard errors. Sur- human capital characteristics.
prisingly, the Breusch-Godfrey test for autocor-
relation (table 5 ) rejects the zero-hypothesis for
no autocorrelation. Since we use cross-section Conclusion
data, the presence of autocorrelation implies
that the functional form is not correct. Because Our examination of the standard Mincer equa-
our research is limited we can not research this tion yields several interesting results. First, using
any further, nor can we perform an Hausman dummy variables for years of schooling seems to
test for endogeneity. Summarizing our findings, better capture the convex relationship between
we obtain the following model: schooling and earnings, as opposed to a contin-
yi 1
uous form. Second, we found that the semi-log
= 0 + 1 xi;educ10 + 2 xi;educ11 + functional form performs better than the logged
3 xi;educ12 + 4 xi;educ13 + 5 xi;educ14 + one. An interesting observation was the presence
6 xi;educ16 + 7 xi;educ18 + 8 xi;educ24 + 9 xi;lexp + of autocorrelation. Since we could not find an ex-
x 2 x 3
10 i;lexp
100 + 11 i;lexp
1000 + 12 xi;f emalelexp + planation for this occurrence, we suggest further
xi;f emalelexp2
13 100 +  i examination. Furthermore, our results suggest
that the human capital characteristics for both
In table 6 we compare our adjusted model with genders differ. To fully capture the difference,
the standard Mincer equation. The coefficients more human capital related data need to be avail-
can be interpreted as an approximate percent- able. Therefore, we suggest further research.
age change of the hourly wage per unit change Although the Mincer equation is not as relevant
of the independent variables. For the effect of as it used to be, we conclude that it still remains
schooling we use several dummy variables, using a valuable equation for research into the relation
the first 9 years of schooling as control group. In of wage and human capital.

2
References
Bjorklund, A. and Kjellstrom, J., 2002, Estimating the return to investments on education: how
useful is the standard Mincer equation?, Economics of Education Review, vol. 21, pp. 195-210.

Card, D. (1999). The causal effect of education on earnings. Handbook of labor economics, 3,
1801-1863.

Edin, P. A., Holmlund, B. (1993). The Swedish Wage Stucture: The Rise and Fall of Solidar-
ity Wage Policy? (No. w4257). National Bureau of Economic Research.

Lemieux, T., 2006, The Mincer Equation thirty years after Schooling, Experience, and Earnings,
Chapter 11 of Jacob Mincer a Pioneer in Modern Labor Eonomics, Grossbard, S. (ed.), Springer.

Norton, A., Cakitaki, B. (2016). Mapping Australian higher education 2016. Grattan Institute.

3
Appendix

4
Table 1: Descriptive Statistics
Variable Obs Mean Std. Dev. Min Max
Hourly wage (y) 5,586 17.142 9.43 2.13 87.5
Hourly wage log 5,586 2.721 0.479 0.756 4.472
Hourly wage box-cox 5,586 2.265 0.329 0.719 3.353
Work experience (lexp) 5,586 22.267 12.475 0 56
Years of schooling 5,586 12.895 2.516 1 24
-9 5,586 .057 0.232 0 1
10 5,586 0.02 0.139 0 1
11 5,586 0.021 0.144 0 1
12 5,586 0.384 0.486 0 1
13 5,586 0.208 0.406 0 1
14 5,586 0.131 0.337 0 1
16 5,586 0.144 0.351 0 1
18 5,586 0.03 0.171 0 1
24 5,586 0.005 0.068 0 1

Table 2: Estimates of box-cox functional form (I) and log functional form (II)

I II
Constant 1.673 (0.021)*** 1.866 (0.029)***
Work experience 0.033 (0.003)*** 0.048 (0.004)***
Work experience squared/100 -0.085 (0.013)*** -0.121 (0.02)***
Work experience cubed/1000 0.007 (0.002)*** 0.01 (0.003)***
Years of schooling
10 0.084 (0.025)*** 0.12 (0.035)***
11 0.092 (0.027)*** 0.133 (0.037)***
12 (High school) 0.239 (0.014)*** 0.342 (0.02)***
13 (College) 0.29 (0.016)*** 0.42 (0.022)***
14 (Associates Degree) 0.399 (0.017)*** 0.576 (0.025)***
16 (Bachelors degree) 0.486 (0.018)*** 0.708 (0.026)***
18 (Masters Degree) 0.832 (0.053)*** 1.221 (0.08)***
24 (Doctorates degree) 0.777 (0.078)*** 1.161 (0.12)***
Work experience 18 -0.011 (0.002)*** -0.016 (0.004)***
Female work experience -0.01 (0.001)*** -0.014 (0.002)***
Female work experience squared/100 0.016 (0.003)*** 0.023 (0.005)***

R-Adjusted 0.239 0.242


Lambda -0.136 0
log-likelihood -944.991 -3028.452
AIC 0.344 1.09
Durbin-Watson 1.81 1.811
Notes: White (1980) heteroskedastic-consistent standard erros; * significant at p < 0.05; ** significant at p
< 0.005; *** significant at p < 0.001.

5
Table 3: Chow Breaktest
H0: Equal coefficients for males and females
F-statistic 0.95 quantile
SSR males 246,843 19,72676 > 6.15
SSR females 209,046
SSR total 476,9163

Parameters 13
Observations 5586

Table 4: Breusch-Pagan test for heteroskedasticity


H0: Constant variance
Variables: fitted values of Box Cox trans. of hourly wage
chi2(1) = 173.14
Prob >chi2 = 0.0000

Table 5: Breusch-Godfrey LM test for autocorrelation


lags(p) chi2 df Prob >chi2
1 51.146 1 0.0000
2 60.700 2 0.0000
3 62.607 3 0.0000
H0: no serial correlation

Table 6: Estimates of box-cox functional form (I) and the standard mincer equation (II)

I II
Constant 1.673 (0.021)*** 1.44 (0.036)***
Years of schooling - 0.073 (0.003)***
Work experience 0.033 (0.003)*** 0.026 (0.002)***
Work experience squared/100 -0.085 (0.013)*** -0.036 (0.004)***
Work experience cubed/1000 0.007 (0.002)*** -
Years of schooling (dummy)
10 0.084 (0.025)*** -
11 0.092 (0.027)*** -
12 0.239 (0.014)*** -
13 0.29 (0.016)*** -
14 0.399 (0.017)*** -
16 0.486 (0.018)*** -
18 0.832 (0.053)*** -
24 0.777 (0.078)*** -
Work experience 18 -0.011 (0.002)*** -
Female work experience -0.01 (0.001)*** -
Female work experience squared/100 0.016 (0.003)*** -

Lambda -0.136 0
log-likelihood -944.991 -3204.244
AIC 0.344 1.149
Durbin-Watson 1.81 1.817
Notes: White (1980) heteroskedastic-consistent standard errors; * significant at p < 0.05; ** significant at p
< 0.005; *** significant at p < 0.001.

You might also like