Introduction
Receiving education is a costly choice. There are costs to education in terms of direct monetary payments and an opportunity costs. The choice to go to school means forgoing waged employment. Therefore, there is an opportunity costs of the wages that would have been earned. Therefore, there is a need to compare the returns versus the cost of education to make optimal decisions. Estimating the returns to education for any population segment is useful in making decisions on the optimal schooling level at both a personal as well as societal level. From a policy standpoint, estimating returns to education will guide decisions on mandatory schooling levels and fiscal allocation to education. This paper seeks to estimate the returns to schooling for Married Women.
Moretti (2004) estimates the returns to education using longitudinal data. The model accounts for unobserved city specific factors using two instrumental variables: city demographic structure, and land-grant for college. The study reveals that a one percent increase in college graduates increases the earnings of high school drop-outs by 1.9 percent, high school graduates by 1.6 percent and college graduates by 0.4 percent. The findings are interesting because the effect is larger for less educated groups which is contrary to expectations.
Dougherty uses National Longitudinal Survey data to assess the returns to education for both men and women. The study reveals that the rate of return for education is for women is much higher than that of men by approximately two percent. The findings are interesting as it is often assumed men earn more than women. The author attributes it to differences in the quality of educational outcome.
Description of Variables
This study was only interested in married women who are already in the labor force. Therefore, the sample size is 3,286.
The study seeks to estimate returns to education. The returns to education are proxied by log of wages per hour. The mean wage per hour is 10.37 with a standard deviation of 7.03. The maximum wage rate is 200 while the lowest was 0.033. The returns are modelled as a function of education. Education was measured as the number of years of schooling. The average years of schooling for the married women were 13.40 with a standard deviation of 2.43. The maximum years of schooling was 18 years and the lowest was 0 years.
There are other factors that influence earnings based on the review of literature. The factors that are identified are experience, union membership and race. Union members are assumed to earn more because the union agitates for higher wages for its members. Race has a bearing on earnings as well.
Experience was measured in years. The mean experience was 19.27 years with a standard deviation of 9.9 years. The maximum experience level was 47 years and the minimum was 0.
Union membership was measured as a dummy variable. 2,795 women were not union members and only 491 were union members.
Race was measured as dummy for Blacks and Hispanics. The sample subset comprised of 199 black women and 183 Hispanic women.
Model Specification
Modelling of expected returns of education can be traced to Becker & Chiswick (1966). Their basic model uses a human capital approach views education as an investment which compares the cost versus the expected discounted future earnings. Therefore, the model argues that total earnings over a person lifetime are the sum of education returns and any intrinsic human capital. Optimal schooling is given as the point where the marginal cost of schooling equals the marginal returns to schooling. The model only considers schooling as the only source of human capital.
Mincer (1975) extended the Becker model by adding experience to the model to account for on job training. Mincer further argues that the relationship experience term is not homoscedastic. They are concave in nature. A Breusch- Pagan test for the present data set confirms that the education and experience are not homoscedastic.
Mincer (1974) presented as estimated the following model
Ln Yi = α + r1 Si + β1ti – β1ti2 + ɛi
The model is widely applied in studies that evaluate the returns to education. However, the OLS estimation has recently been criticized because it ignores individual heterogeneity of earnings. Besides, the OLS estimates are seen as bias because the correlation between unobserved ability and education level which creates an ability bias.
Instrumental Variable regression has been applied to account for the influence unobserved individual ability differences on education choice. Some of the applied instruments include tuition costs, distance to college and minimum education laws. It was not considered because a measure of ability could not be included in the model.
Given limitations of the dataset, there is no precedent for an instrument to use. This study hypothesizes that there is a correlation between husband education and wife education. This is because people tend to marry from their social circles. Therefore, people spend their youthful along those they are studying with and are likely to select a partner from the pool. Besides, educated husbands are likely to encourage their wives to advance education compared to less educated husbands. However, there is no direct link between husband education and ability. The correlation of husband education and wife education is confirmed by the scatter plot. An OLS regression confirms the correlation between husband education and wife education.
An IV regression is applied on the model developed by Mincer which is augmented by additional variables. Dummy variables for ethnicity and union membership are added to the model. Therefore, the model that is estimated is as follows using an IV regression 2 stage least square with husband education as the instrumental variable;
Structural equation
Ln Wage = α + r1 Si + β1ti – β2ti2 + + β3 U + β4 B + β5H + ɛi
Where
S is the years of schooling
t is the experience in years
U is dummy for union membership (1 if one is union member)
B is dummy for Black (1 if wife is black)
H is dummy for Hispanic (1 if wife is Hispanic)
Reduced Equation
S = π0 + π1Zi + υi
Where
Z is the years of schooling of husband.
Results and Interpretation
The model is statistically significant according to the Wald test. Therefore, all the regressors are jointly significant determinants of the log of wages. The R-square for the model is 0.1955. Therefore, the model explains atleast 20 percent of variation in the log of wage. This means that there are other factors that explain 80 percent of the variations in wages that were not included in the model due to data limitations. It is plausible that it maybe a source of omitted variable bias. Including the variable with improve accuracy of the size effect of education that we observe.
The coefficient for education is 0.131. There is a positive relationship between education and hourly wages. An increase in the years of schooling by one year increases the hourly wage by 13.1 percent. Education is a statistically significant determinant of hourly wages.
The coefficient for exprience is 0.0175. There is a positive relationship between exprience and hourly wages. An increase in the years of schooling by one year increases the hourly wage by 1.75 percent. Experience is a statistically significant determinant of wages.
The coefficient of experience squared is negative 0.00024. The co-efficient is statistically signifcant. It confirms the earlier asumption that the marginal returns to experience are not constant. Experience assumes a concave shape.
The coefficient for union membership is 0.144. The coefficient is statistically signifcant. This means that union members earn more than none union members.
The coefficient for black women is -0.029. The coefficient is statistically significant. This means that black women earn relatively less compared to the other ethnic groups.
The co-efficient for hispanic women is 0.072. The coefficient is statistically significant. This means that hispanic women earn relatively higher than other ethnic groups.
The IV regression assumes that education is an endogenous variable. We tested the assumption using Durbin and Wu-Hausman test.
P-value = 0.000 < 0.05 for both Durbin and Wu-Hausman tests. From the Durbin test and Wu-Hausman test, we reject the null hypothesis that the variable is endogenous at 5 percent significance level. Therefore, education is endogenous and not exogenous as assumed in the model. We observe a positive relationship between the log of wage and education. This means that highly educated women earn more than less educated women. Therefore, married women should enhance their education to improve their earnings.
We used husband education as a instrumental variable for the IV regression. Therefore, we assumed that husband education is a valid instruemental variable. We use the Wald test to test the assumption.
The partial R-square shows the correlation of education and husband education after controlling for the other instruments. It is moderate. P-value = 0.000 < 0.05. We also reject the null hypothesis that the instrument is weak at 5 percent significance level.
Therefore, the two assumption hold. Education of married women is an endogenous variable and the husband education is a valid instrument for the model. This allows us to partially counter the ability bias from the model.
Diagnostic and Sensitivity Analysis
We use an OLS model to estimate the returns to schooling as a robustness test.
The model has an adjusted R-square of 0.2161. This means that the model only explains 21.61 of the factors that influence the wages of married women who are in the labor force. There are other factors which explains approximately 80 percent of the differences in hourly wage rate. The most obvious factor that is ignored by the present model is ability. The limitation of the dataset make it impossible to include a proxy of ability. Previous studies have proxied ability using the results of standardized IQ tests.
The model has a F-statistic of 153.69 with a p-value of 0.000. Therefore, all the factors are jointly statistically significant at 5.
The coefficient for education is 0.0947. There is a positive relationship between education and hourly wages. An increase in the years of schooling by one year increases the hourly wage by 9.47 percent. Education is a statistically significant determinant of hourly wages. The size effect is lower compared to IV regression.
The coefficient for exprience is 0.0173. There is a positive relationship between exprience and hourly wages. An increase in the years of schooling by one year increases the hourly wage by 1.73 percent. Experience is a statistically significant determinant of wages. The coefficient is close to that of IV regression.
The coefficient for union membership is 0.181. The coefficient is statistically signifcant. This means that union members earn more than none union members.
The coefficient for black women is -0.046. The coefficient is statistically significant. This means that black women earn relatively less compared to the other ethnic groups.
The co-efficient for hispanic women is 0.002. The coefficient is statistically significant. This means that hispanic women earn relatively higher than other ethnic groups.
There are various potential violations of the model that could be not be corrected due to restirctions of the data. Firstly, as earlier mentioned, omitted variable bias is likely because ability is not included in the model. Individuals have intristic human capital that is augmented by schooling and job training. Therefore, ignoring ability is likely to result in higher effects for education and experience that it actually it. Ignoring differences in ability ignores the hetoerogenity expected in any population. Besides, there is a correlation between ability and education. The marginal cost of education is expected to be lower for those high-ability persons compared to low ability persons. High ability persons will take less time at school and are more likely to receive scholarships. Therefore, they are more likely to advance to higher education levels. There insturmental variable partly corrects it. However, the instumental variable selected is based on the resaercher judgement. If more data was available, previously used insturmental variable could be applied.
Education level and experience are likely to be misreported since the data is suvey data. Self-reported schooling and experience data are often not accurate. This creates a measurement error creating an attentuation bias. Given that the returns to experience and education are positive, the bias is downwards.
Conclusion
The study uses an IV estimation to estimate the marginal returns of education of married women already in the job market using husband education as an instrumental variable. The study reveals that an increase in one year of schooling creates an increase of hourly wages by 13.1 percent. However, the study does not include ability. If more information was available, a proxy for ability could be included in the model. Analysis of returns to education for husbands who are in the job market is necessary to allow for comparison across gender. The study also acknowledges that there are non-monetary returns to education. It is impractical to quantify those returns. Therefore, this study ignored non-monetary returns.
Bibliography
Becker, G. & Chiswick, B., 1966. Education and the Distribution of Earnings. American Economic Review, pp. 358-369.
Dougherty, C., 2003. Why is the Rate of Return to Schooling is Higher for Women than for Men. Center For Economic Performance.
Mincer , J., 1975. Education, Experience and Distribution of Earnings: An Overview. Education, Income, and Human Behavior.
Moretti , E., 2004. Estimating the social return to higher education:. Journal of Econometrics evidence from longitudinal and repeated cross-sectional data, pp. 175-212.