WHAT DETERMINES WAGES? ESTIMATING THE IMPACT OF EDUCATION ON WAGES IN THE UK. EVIDENCE FROM THE 2002 LABOUR FORCE SURVEY TEACHING DATASET
Abstract: The economics of education is at the root of human capital theory which is what forms the backbone of this study. OLS procedures have been used to calculate the wage premium by educational qualifications using cross sectional data from the 2002 LFS Teaching Dataset. This study also contains a brief section that uses quantile regression analysis and compares the wage premium estimates with OLS.
Section 1: Introduction
This study estimates the wage premium earned for attaining various academic qualifications under the basic human capital earnings equation and then examines how these estimates change after controlling for other characteristics such as gender, sector of work ethnicity and other variables. Having no suitable measure for ability or family background is bound to cause our estimates to be biased, but this study attempts to reduce the bias as much as possible by adding relevant control variables to explain the variation in (log) hourly earnings. Controls for gender are interacted with education and marital status for qualitative information about earnings by gender, marital status and education. Highest qualification has been used as the measure for education for all specifications, but some results have also been estimated using (potential) years of schooling for the sake of comparison with previous studies. Our estimates for returns to education using years of schooling fall within the 7-10% band within which is what most studies for the UK have found. Since OLS is violated due to heteroskedasticity, a basic quantile regression analysis is run in Section 5 as is does not need to satisfy the assumption of homoscedasticity. We find tentative evidence that women appear to be subject to a greater unequal treatment in higher paid jobs than in low paid jobs. We also find that the wage premium rewarded for obtaining higher academic qualifications is higher for the top 10% of the conditional wage distribution than for the bottom 10% of the conditional wage distribution.
Section 2: Literature Review
Returns to education are the private gains that individuals make by investing in their education. The human capital approach follows that education is an investment of current resources in exchange of future returns. Jones (1993) likens an investment in education to an investment in a machine which can be fitted on to the human body to enhance one’s performance in the workplace. If the returns to such a machine or educated person exceed the outlay of time and money involved in buying it, it is a good investment.
It is hard to unravel the pure effects time spent in education and the qualifications achieved on earnings which is why evaluating the rate of return to education is not facile. The econometric and theoretical advances in measuring the causal effect of education on labour market earnings are well explained by Card (1999). Education has been measured as years of schooling in the USA, but by highest qualification attained in the UK. Card (1999) justifies this on grounds that graduating from high school in the UK may entail different years of schooling depending on whether a student plans to pursue higher education or work straight away.
The ‘years of schooling’ measure assumes that every year of education adds exactly the same percentage amount to wages regardless of the specific year of schooling. This seems far-fetched but there is insufficient evidence to conclusively disprove this hypothesis (see Card 1999; Harmon et al. 2000). Plotting average return for qualifications against average years of schooling and fitting a simple regression line through these points clearly shows a linear form for the UK (using BHPS1). This finding by Harmon et al. (2000) led them to deduce that while linearity is a strong assumption, it is quite difficult to invalidate. In sharp contrast, Walker and Zhu (2002) soundly reject the assumption of linearity for women and marginally for men using data from Labour Force Surveys (1992-2000). After allowing for each year of education to have an independent effect on earnings, they find that additional returns are lower at higher levels of education. However, the approximate linearity of earnings with respect to schooling is a useful simplification because it can aid the formulation of tractable econometric models.
Ability bias is a key issue in the literature of the economics of education even though some economists hold the view that it is probably overstated by critics of the human capital paradigm (see Becker 1964, cited in Card 1999). Ability is an essential but unobservable component of human capital that determines wages. Omitting it biases our estimates for the return to education upwards since more educated workers also have higher levels of unobserved ability. This can be dealt with by including a suitable proxy for ability in the earnings equation. Fallon and Verry (1988) explain that the proxy used must refer to a point in time prior to the acquisition of schooling since any measure of current ability will pick up the ability raising effects of education and partly conceal returns to the latter. However, it is unclear whether IQ tests or other measures such as the AFQT2 provide adequate proxies (see Griliches 1977; Fallon and Verry 1988).
Another way to deal with ability bias is to treat schooling symmetrically with ability measures, allowing it to be subject to measurement error and to be correlated to the error term in the Mincerian equation. This way, the implied net bias would be zero or negative (Griliches 1977). The theoretical prediction that omitted ability and measurement error cancel each other out (upward ability bias and downward bias via measurement error) is confirmed by the findings of a UK based twin study by Bonjour et al. (2003). In contrast, a US based twin study by Ashenfelter and Krueger (1994) found that omitting ability does not cause upward bias, but measurement error does cause downward bias. As opposed to Bonjour et al., this (2003) implies that ability bias and measurement error do not cancel out.
It is plausible to think that ability by itself would not necessarily bias least squares estimates upwards. In line with this view, Walker and Zhu (2002) conjecture that there could be unobservables other than ability which impact schooling (S) and wages (w) in opposite directions. This would bias OLS estimates downwards. They explain that impatient individuals have a high time preference because they lay more emphasis than average on immediate well-being compared to future well-being. Since they discount future gains more heavily, they have a high discount rate3. Obtaining less years of education would be most favourable for such individuals. However, if employers believe that impatient individuals have the drive to meet deadlines and complete tasks quicker than others, then they might be rewarded for having this trait. In other words, impatient individuals may have a high opportunity cost (foregone labour market rewards) which reduces r. Hence the answer to what kind of bias OLS estimates have is ultimately an empirical matter.
The literature treats the schooling measure as exogenous, although it is well established that education is an endogenously determined. The research that deals with endogeneity of schooling can be classified into three groups. The first involves finding a suitable proxy for ability as we have already discussed earlier. The second exploits within twins differences in wages and education. Estimates obtained from twin studies (for example: Ashenfelter and Krueger 1994, Bonjour et al. 2003) and other natural experiments are considered to be the most reliable of all. However, Walker and Zhu warn that (2002) this requires us to accept the assumption that that unobserved fixed effects are additive so that they can be differenced out by regressing the wage difference within twins against the education difference. The third method involves the use of instrumental variables which could provide more accurate estimates than ordinary least squares (OLS) if they meet two conditions: (i) being uncorrelated with the error term and (ii) being partially correlated with the independent variable (education). If not, it is might be better to use OLS estimates. In practice, even though it is very difficult to find a suitable instrument variable (IV), recent literature has turned to the usage of institutional reforms as IVs. Their usage is justified on the grounds that it is close to a natural experiment.
That said, it is noteworthy that the estimation method does not necessarily make a big difference on the returns to education. Some findings suggest that the average marginal return to education in a given population is just slightly below the estimate that emerges from a simple cross-sectional regression of earnings on education (see Psacharopoulos and Patrinos 2004, Card 1999; Griliches 1977). Angrist and Krueger (1991) found that their IV estimate and Wald estimate for the return to education was similar to OLS estimates suggesting that there is little bias with conventional OLS estimates.
Finally, in line with the views of Psacharopoulos and Patrinos (2004), a major research gap in the existing can be seen in the inconsistency between micro and macro evidence on the returns to investment in education. While it is well established that there are measurable returns to education in the micro case, there is no such evidence in the macro literature.
USA based studies predict that the returns to education is 10%. Evidence5 from the UK shows that this figure ranges between 7-9%. UK has one of the highest returns to education in comparison to other European countries. Scandinavia, in particular has low returns to education. Psacharopoulos and Patrinos (2004) consider twin studies and other natural experiments to be the most reliable for estimating the returns to education. However, such studies are difficult to carry out, let alone track over time. Bonjour et al.(2003) is perhaps the only study (in the UK) conducted using a sample of over 3000 twin pairs. However, this also suffered from data limitations and was conducted only for female twins.
Section3: Methodology
This study uses Ordinary Least Squares (OLS) estimates to analyse the Mincerian earnings function which specifies log earnings as a simple function of highest educational qualification and experience. This can be augmented and enhanced to inform on an array of policy issues other than those related to human capital. The estimation of more elaborate equations has allowed the estimation of the size of gender pay gaps, regional wage differentials and wage structures by sector of work. Many of the insights obtained about wage inequality are based on the estimation of wage equations.
The standard Mincerian earnings equation is of the form:
ln(wageπ)=πΌ+πππ»ππ+π1EXPERπ+π2EXPERπ2+π§i ,where zi=other variables [1]
Our estimated model is of the form:
ln(wageπ)=πΌ+π1π»ππ+π2EXPERπ+π3EXPERπ2+π½MALEπ+xiΞ³ +π’π [1.1]
Where i=1,2…,n are the cross-sectional data points (for individuals) and ui is the error term. The variable HQi represents an individual’s highest qualification, while EXPER (and its quadratic) capture an individual’s experience in the labour market. The coefficient on HQi can be thought of as the premium (private) wage return to a certain level of education. Note that X (experience) is measured by differencing an individual’s age and the age at which schooling has been completed (6 years); assuming that individuals start working as they finish schooling. MALEi=1 if the individual is male and zero if female, and xi is a vector comprising an individual’s employment status, ethnicity, marital status, sector of work and other relevant controls.
Under standard textbook assumptions, Ordinary Least Squares (OLS) can be applied to equation [1]. The assumptions are contained in: ui ~N(0,Ο2). It is reasonable to assume that labour market earnings follow a log-normal distribution. We also assume a constant error variance in the log earnings equation. We expect the assumption of normality to be violated since OLS estimates are basically sample average statistics are sensitive to the presence of outliers. Quantile regressions are effective procedures if outliers are considered the major source of the problem (Reilly 2013). OLS is a mean (on average) procedure and the resultant estimates are just numerical values that capture the relationship between the independent and dependent variable. Table 3 offers a comparison between OLS and quantile regression estimates for our data along with a simple interpretation of the results. The quantile regression relaxes the assumption of homogeneity across the conditional distribution. It is assumed that regardless of where you are on the conditional log wage distribution, the effect of gender or education on log wage is assumed the same. A detailed analysis of quantile regressions is beyond the scope of this study. The main focus of this study is estimate the determinants of (log) hourly wages and to discuss how earnings vary by controlling for a variety of characteristics.
The advantage of estimating the rate of return by the dummy variable method rather than the years-of-schooling-squared method is that a great deal of sensitivity is added to our estimation and the rate of return may not be as sensitive as suggested by the standard formula. However, the problem with the earnings function approach in general is that the rates of return are estimated on three implicit assumptions. First, the age earnings profiles are either flat between adjacent educational levels throughout their range. Second, the age earnings profiles last forever. Thirdly, the only cost of schooling is the foregone earnings of the individual. These assumptions are difficult to satisfy but are essential for making reasonable interpretations of our data.
Section4: Data Sources and Description:
This study uses cross-sectional data from the Labour Force Survey 2002 Teaching dataset6 which contains information from all four quarters of the 2002/3 LFS Teaching Dataset for respondents aged 16-65 and residents in the UK (n=63,559). This is a subset of the original 2002 Labour Force survey modified for student and teacher use. Sampling procedures have been followed to obtain a random sample via telephone interview, but weighting has not been used. Filtering the data by excluding negative values for hourly wages (originally coded to represent responses such as ‘don’t know’ and ‘does not apply’ when asked about hourly earnings) lead to the reduction to our sample size by one half. Further transformations excluded observations for individuals that were not within the official working age in 2002 (women:16-59 years; men: 16-64 years). Observations that reported ‘don’t know’ or ‘other qualifications’ when asked about highest qualification obtained were also dropped from our sample. Our final sample size fit for regression analysis contains 25512 observations.
The summary statistics reveal that there is nothing untoward about the data since all variables report 25512 observations. In addition, the dummy variables are coded with either 0 or 1 and if we add the proportions for the mutually exclusive dummy variables, they add to one. There are no implausible negative values reported for any of the explanatory variables and the zero values for experience possibly relate to people that just started working when they were interviewed. Our main measure of education is the highest qualification attained by individuals captured by five dummy variables. Note that all higher qualifications above A-level but below degree level are categorised under the dummy variable ‘Diploma’. The remaining binary dummies for education: ‘Degree’, ‘A-level’ and ‘GCSE’ also tell whether respondents attained academic or vocational equivalents of these qualifications (e.g. GCSE contains information about attained an ‘O-level’ pass and NVQ17). The second measure of education: ‘school’ denotes potential years of schooling (age finished schooling- 6 years). This is used to obtain only a few results because highest qualification is a more appropriate measure of education for the UK. The average (potential) years of schooling is almost 12 years and almost 47% of the sample is married.
About 60% of the respondents in our sample are reportedly married, almost one-third are single and the remaining are divorced. A kernel density plot for log wage shows that the distribution of log wages is asymmetric and skewed to the right (see fig2.a). This is confirmed by the Jarque-Bera normality test.
The p-value is 0 and the chi-squared is so large that it is not even reported in the STATA output. The test confirms that the two underlying propositions governing normality (i.e. symmetry and meso-kurtosis) are decisively rejected in this case. The density plot suggests that there are more highly paid workers in the log wage distribution than predicted by a normal distribution.
Our dataset shows that 33% of all ethnic minorities live in South-East England, which is the most prosperous region of the UK (see Fig2e.,appendix). People from the ‘Mixed ethnic group’ labelled as ‘other’ formed the largest proportion of ethnic minorities in the UK at 27% followed by the Indian community that forms 22% of all ethnic minorities. Other than the White British group, those most likely to be born in the UK were individuals from the Mixed ethnic group and from the ‘Other Black group’. Figure 2c. shows that about half of the non-white population in the UK comprises of Asians. Blacks form almost one-third of all ethnic minorities. Individuals from mixed or other ethnic backgrounds form one-quarter of the non-white population in the UK. Indians form the largest group from Asia within the UK followed by Pakistanis.
Men and women with higher academic qualifications have higher average hourly wages than individuals with lower academic qualifications (Fig2d). However, females earn a lower average hourly wage at every level of education than males in the UK labour market. This earnings gap shall be examined in further sections.
An important caveat is that the teaching dataset has been prepared solely for the purpose of teaching and student use. Some variables have been recorded using simplified procedures and may vary from the original LFS data. This might lead to inaccuracies and might bias our results as there is no information about how exactly variables have been modified from the original LFS 2002/3. In addition, there is a possibility of attenuation bias if we are willing to believe that some individuals might give incorrect information (intentionally or unintentionally) about hourly wages in telephone interviews. Due to the nature of earnings distribution, we are quite likely to encounter heteroskedasticity, which can be dealt with by using robust standard errors.
Section5: Model Estimation and Hypothesis Testing
This study begins with an ordinary least squares specification and finds that the education variable defined by ‘highest qualification’ explains a significant part of the variation in log earnings. While the wage equation explicitly shows that we want to hold ability fixed while measuring the return to education, there is no suitable variable that can act as a proxy for education. The constant is excluded from all specifications because we do not really care whether we get an unbiased or consistent estimator of the intercept Ξ²o as this is usually not possible.
Log (wage) = Ξ²o + Ξ²iHIQUALDi + u [1]
By simply regressing log (wages) on the educational dummies, OLS explains about one-quarter of the variation in earnings. Adding experience and its quadratic gives us the Mincerian earnings function which explains over one-third of the variation in earnings (see Table 1 on p.14 column (1) and (2)). All specifications suffer from omitted variable bias as we have no information about two important variables that determine earnings i.e. ability and family background. Neither does our dataset provide any suitable proxies for ability or family background and nor does it provide any variables that satisfy the two conditions for a using a valid IV. We cannot really hope to ever calculate the coefficient for ability because ability is a rather vague concept at best. Even so, we can reduce the bias term by controlling for a number of important variables that are available in the LFS 2002 Teaching dataset.
Table1. shows six OLS regression models that control for different characteristics. It is observed that the premium to attaining a given level of education is quite robust despite controlling for a number of other variables. In most specifications, individuals with a University degree earn over 80% more than individuals with no qualifications, ceteris paribus. Likewise, on average and ceteris paribus, individuals that possess an A-level qualification earn over 30% more than individuals with no qualifications.
It is evident from column (4) and (5) that on average a public sector worker earns 2.5% [(e0.0243-1)x100] more than a private sector worker with the same level of education, experience and other characteristics. After controlling for job status (i.e. permanent or temporary), we find that public sector earnings are almost 4% higher than private sector earnings for individuals with identical characteristics ceteris paribus.
Only three (white, Black and Indian) of nine dummies for ethnicity yield statistically significant results for column (5) at 5% level of significance. Column (5) suggests that on average and ceteris paribus, whites earn over 7% more than Pakistanis whereas Indians earn over 10% more than Pakistanis working in the UK. Workers born outside the UK earn roughly 0.47% more than individuals born in the UK. There is no strong reason to believe why this might be true, but this could be the case since 7% of our sample is born outside the UK, of which 61% of them work in South-East England which the most prosperous region in the UK. Individuals working in this region earn about 27% more than those working in Northern Ireland, ceteris paribus. Likewise, holding all else fixed, workers in Scotland earn 9% more than those in Northern Ireland.
Column (6) tells us that on average and ceteris paribus, an additional year of experience leads to a 3.7% [{0.0389+ 2x (-7.13x10-4)}x100] increase in hourly earnings. Male earnings are 26% higher than female earnings [[(e0.231-1)x100] given that both have the same level of education and marital status. After controlling for part-time work status and sector of work in column8 (4), we see that the gender-wage gap decreases by almost 9 percentage points. Columns (4-6) suggest that on average and ceteris paribus, females earn around 17% less than males.
A part-time worker earns approximately 24% less than an individual with the same level of education, gender, sector of work and other characteristics. This result is well estimated and is robust to changes in ethnicity and region. We also find that the effect of having a permanent job are well determined. The point estimate 0.085 suggests that holding all else constant workers with permanent jobs earn 8.9% [(e0.085-1)x100] more than individuals with temporary jobs. A simple regression of earnings on‘permanent job’ shows that 0.2% of the variation in earnings is explained by whether an individual has a permanent or temporary job. Omission of job status has contributed to the upward bias in our OLS estimates (apart from the upward bias caused by omission of ability). For this reason, the coefficients for all educational dummies in column (6) have fallen in comparison to all the previous columns.
It would be even better to be able to calculate the returns to education individually for each level of education, however, this would require precise knowledge of how long a ‘Diploma’ takes to complete. Note that this dummy variable covers a range of qualifications (higher than A-levels and lower than a Degree) which take different years to complete. Due to this ambiguity, we resort to simply calculating the premium returns to obtaining highest qualification with respect to its preceding qualification. This can be calculated by simply differencing the coefficient of the succeeding and preceding qualification. By this simple calculation for column (6) we estimate that the wage premium of acquiring a university degree compared to just an A-levels qualification is almost 50 %. Likewise, the wage premium of having passed A-levels compared to just having qualified GCSE examinations is 10.7%. We are more interested in the premium earned by completing a degree vis-Γ -vis acquiring A-levels rather than a 1-2 year diploma for two reasons. First, almost one-third of all individuals in our sample have reported ‘A-level’ as their highest qualification compared to just one-tenth of our sample that has completed a diploma or its equivalent. Second, acquiring ‘A levels’ qualification is the gold standard entry requirement for getting a university degree.
That said, it is reasonable to assume that a university Degree takes three years to complete after completing A-levels. As a rough estimate from column (6), we could say that the returns to obtaining a university degree over an A-levels qualification is 16% [(0.797-0.307)/3]. Likewise, we could say that the returns to obtaining an A-levels qualification is 5.3% [(0.307-0.2)/2] given that it takes two years to complete an A-levels qualification after GCSE.
For the sake of comparison with previous studies and to get an estimate for the returns to an additional year of schooling, the years of schooling measure has been used to get rough estimates for returns to schooling (appendix, table4). Table 4 has exactly the same specifications as Table1, except that education is measured as years of schooling. The final specification shows that an additional year of education leads to a 9% rise in (log) hourly earnings, ceteris paribus.
While this is consistent with most findings in the literature which suggest that returns to an additional year in education are between 7-10% our estimates are not reliable since we ought to be using highest qualification as a measure for education for the UK. Table 4.a. uses the same specifications as in column (6) in Table 4 for males and females separately. It suggest that the returns to an additional year of schooling are higher 0.23 percentage points for females than for males even though males hourly earnings are higher than females. Figure 1 plots these rates of return for men and women. As expected we see that returns to education are higher for women than for men.
We can recast the models developed in Table 1 earlier by adding an interaction term between marital status and gender to the model where female and married appear separately. This allows the marriage premium to depend on gender. Table 2 helps us understand how male/female earnings vary by level of education and marital status. We find a statistically significant relationship between gender and marital status in columns (3) to (5). Column (3) shows us that on average and ceteris paribus, married males earn 22.5% more [{0.189-(-0.0356)}x100] than single males. Upon inclusion of part-time work status in column (4), we see that this gap decreases to 12.38%. Column (3) shows that married females earn 3.56% less than single females, ceteris paribus. This makes intuitive sense as many women work part-time or stop working altogether after marriage. The point estimate for part-time work status is -0.2. This is not surprising as part time jobs are known to be less prestigious (than full-time jobs) and have a low skill requirement.
Interacting males with highest qualification in column 5 allows the gender premium to depend on education. We see that there is no statistically significant wage differential between males and females that hold the same qualification having controlled for part-time work status, sector of work and other characteristics. There is no evidence against the hypothesis that the premium to education is the same for men and women with comparable levels of education. However, we cannot conclude that there is no statistically significant evidence of lower pay for women at the same levels of education, experience and other characteristics. This is because we have added the interaction male.educi to the equation and the coefficient on male is estimated much less precisely than it was in Table1. The standard error for male has increased by over two-fold from model (3) to (5) after controlling for the interactions of the educational dummies with ‘male’.
Quantile Regression-a brief mention and simple comparison with OLS model
An alternative methodology to OLS is available known as quantile regression (QR) which allows the estimation of returns to education within different quantiles of the wage distribution based on the entire sample (Buchinsky 1994, cited in Harmon et.al. 2000). This requires a sufficiently wide spread of education for which our data seems to be satisfactory. Table 3 shows that the coefficients for highest qualification are statistically significant for each decile. The premium earnings for obtaining a degree in the top 10% of the conditional pay distribution is significantly higher than the premium earnings to obtaining a degree in the bottom 10% of the conditional pay distribution.
Table 3. summarises the estimates obtained for various regression models using the LFS 2002 dataset. The estimated gender effect is well determined and is statistically significant for the mean regression model and for all the models estimated at different quantiles of the conditional pay distribution. OLS regression estimates in column (1) show that ceteris paribus men earn almost 18% more than women [(e0.163-1)x100] which is quite close to the median regression estimate which suggests that men earn almost 17% more than women holding all else fixed. Holding all other factors fixed, at the top 10% of the conditional wage distribution, males earn 22% more than females whereas at the bottom 10% of the conditional wage distribution, males earn only 13% more than females. In other words, we observe an increase in the gender differential as we move along the conditional pay distribution suggesting that women experience a ‘glass ceiling’ compared to men in the UK labour market.
The unequal pay treatment is higher at higher paid than lower paid jobs. This ‘glass ceiling’ effect is also supported by the inter-quantile regression estimates which show that there is a ceteris paribus increase of almost 8% in the pay gap between men and women between the 90th and 10th percentiles. This is computed as (e0.0736-1)x100. Note that all the coefficients for male are statistically significant.
Also note that the median point estimate for male is 0.154 which is lower than the mean regression estimate (0.163) suggesting that the log (wage) distribution is sensitive to outliers. Figure (2a.) gives tenuous evidence for this. It may well be that many of the individuals in the elongated right tail are actually men. The wage premium for highest qualification can be calculated as follows:
Wage premium=(coefficient for succeeding qualification-coefficient for preceding qualification) x100 %
SECTION 6: LIMITATIONS
This study has ignored the effect of fixed characteristics such as family background, motivation, school quality and ability on the return to education. There is a need for more research to understand the extent to which family background affects an individual’s highest achieved qualification.
LFS data is known to be reliable as rigorous sampling techniques are used to cover the population as a whole, but data quality still causes some difficulties in precisely estimating the returns to education. Adding controls for hours worked in main job as well as hours worked part time would have led us to get more reliable OLS estimates. However, while we have information about hours worked in main job, we have no information on the number of hours worked part-time.
It is difficult to say whether a simple relation between earnings and education can be interpreted as a return to education for a randomly selected person. To support such an interpretation, one must convincingly control for factors such as ability and family background that might both affect the choice of education and wage. We do not have any measure of ability or family background, both of which are important determinants of earnings. Hence, this study has serious potential credibility problems. One must render caution in interpreting causality due to endogeneity. An additional caveat is that the while we assume for simplicity that the zero conditional mean assumption holds in our study, it is in fact a very difficult condition to satisfy. In fact, there is almost certainly quite substantial correlation between the error term and unobserved omitted variables such as ability and motivation.
The main difficulty with LFS 2002 Teaching dataset arises from the censored nature of the data on age at which the individual left full-time education. The LFS question refers to continuous education so that interrupted spells of education may be censored. In particular, a “gap” year between secondary and higher education could result in the spell of higher education not being reported. However this is a relatively minor issue in comparison to ability bias10Also, OLS violates the Gauss Markov assumptions and OLS is no longer the best linear unbiased estimator.
This study uses OLS regression analysis to estimate the effect of education on earnings. Since we do not have reliable measures for years of schooling, which is the conventional measure for education, we resort to using ‘highest qualification’ attained by an individual for our analysis. The basic assumptions for OLS especially the ones contained in the zero conditional mean assumption are violated. Education which is treated exogenous in the literature is in fact endogenous and this leads to bias.
The results in this study are not suited for direct application and policy framing due to the lack of reliability that stems from bias. Careful extrapolations made from a conceptual standpoint, however, are likely to produce more efficient workable solutions. One could argue about the possibility that measurement error and ability bias might cancel out giving us unbiased OLS estimates, however, that would be a naive conjecture since the extent of ability and measurement error is hard to determine in the first place. Since we have heteroskedasticity, robust standard errors have been included in all OLS specifications. Also note that ‘Years of Education’ in the LFS can only be inferred from the age at which individuals left full-time continuous education. Thus, the data fails to record years of education accurately for those that interrupted their schooling.
SECTION 7.CONCLUSION
A simple regression model with educational dummies and a low order polynomial in potential earnings explains about one-third of the variation in observed earnings data. This model is too parsimonious to fully characterise the distribution of earnings and schooling. However, it does serve as a convenient benchmark from which more complex models of (log) earnings can be determined by examining the effects of covariates such as ethnicity, gender, sector of work and other individual specific characteristics.
This study uses OLS regression analysis to estimate the effect of education on earnings. Since we do not have reliable measures for years of schooling, which is the conventional measure for education, we resort to using ‘highest qualification’ attained by an individual for our analysis. The basic assumptions for OLS especially the ones contained in the zero conditional mean assumption are violated. Education which is treated exogenous in the literature is in fact endogenous and this leads to bias. Using quantile regressions, we find evidence that women experience glass ceiling compared to men in the UK labour market. We also find that an individual in the top 10% of the conditional pay distribution is rewarded more for having the same educational qualification as another individual in the bottom 10% of the conditional pay distribution, ceteris paribus. We estimate that the wage premium to obtaining a Degree over having an A-levels qualification is almost 50% and the premium of attaining an A-level qualification over GCSE is around 10%.
The results highlighted in this report treat education as exogenous. While it is difficult to deal with this problem effectively in the LFS data, it should be noted that majority of previous research suggests that doing so would lead to higher estimated returns. Walker and Zhu (2002) have explained that since most reforms have aimed at increasing educational participation amongst those most likely to leave school at an early age, it seems unlikely that the IV results in the literature are relevant to low education individuals. However, whether quantile regression methods would replicate findings of higher marginal returns remains to be seen. This could be an avenue for future research.
Note*: Tables have been omitted for readability
References
Angrist, J. D., & Krueger, A. B. (1991) ‘Does compulsory school attendance affect schooling and earnings?’ The Quarterly Journal of Economics, 106(4), pp. 979-1014.
Ashenfelter, O. & Krueger, A. B. (1994), ‘Estimates of the economic return to schooling from anew sample of twins’, American Economic Review, 84(5), pp. 1157–1173 (Accessed: 22 April 2014)
Blackaby, D. H., Leslie, D. G., Murphy, P. D., & O'Leary, N. C. (2002) ‘White/ethnic minority earnings and employment differentials in Britain: evidence from the LFS’, Oxford Economic Papers, 54(2), pp. 270-297.
Bonjour, D., Cherkas, L. F., Haskel, J. E., Hawkes, D. D., & Spector, T. D. (2003), ‘Returns to education: Evidence from UK twins’, American Economic Review, pp. 1799-1812, JSTOR [Online]. Available at: www.jstor.org/stable/3132153 (Accessed:09 April 2014)
Bound, J., Jaeger, D. A., & Baker, R. M. (1995), ‘Problems with instrumental variables estimation when the correlation between the instruments and the endogenous explanatory variable is weak’., Journal of the American statistical association, 90(430), pp. 443-450.
Card, D. (1999). The causal effect of education on earnings. Handbook of labor economics, 3, pp.1801-1863. Available at: http://dx.doi.org/10.1016/S1573-4463(99)03011-4 (Accessed:09 April 2014)
Fallon, P. and Donald, V. (1988) ‘Earnings and earning differentials’ The Economics of Labour Markets. Hertfordshire: Philip Allan, pp. 134-163
Griliches, Z. (1977), ‘Estimating the returns to schooling: Some econometric problems’. Econometrica: Journal of the Econometric Society, 1-22. JSTOR [Online]. Available at: http://www.jstor.org/stable/1913285 (Accessed: 22 April 2014)
Harmon, C., & Walker, I. (1995), ‘Estimates of the economic return to schooling for the United Kingdom’ ,The American Economic Review, 85(5), 1278-1286.
Harmon, C., Oosterbeek, H., & Walker, I. (2000). The returns to education: a review of evidence, issues and deficiencies in the literature, Centre for the Economics of Education, 85491
34
London School of Economics and Political Science. Available from : http://eprints.lse.ac.uk/19551/1/CEEDP05.pdf
Jones, G (1993) The Economics of Education. London: The Macmillan Press Ltd.
LSE Research Laboratory Data Service [distributor]. Qualifications in the United Kingdom 1985–1999 . Available from: http://rlab.lse.ac.uk/data/depository/terms.htm (accessed 2 May 2014)
Office for National Statistics (2005) 2001 Census: Ethnicity and Identity in the UK, Summary Report, [Online] Available from: www.statistics.gov.uk [Accessed 25 April 2014]
Office for National Statistics (2014) Public and Private Sector Earnings-March 2014, [Online] Available from: www.ons.gov.uk/ons/rel/lmac/public-and-private-sector-earnings/march-2014/rpt---march-2014.html [Accessed 17 April 2014]
Psacharopoulos, G., and Patrinos, H.A. (2004) ‘Returns to investment in education: a further update’, Education Economics, 12(2), pp.111-134, DOI:10.1080/0964529042000239140 (Accessed: 09 April 2014)
Reilly, B. (2013), Lecture One Notes: OLS Wage Equation and Quantile Regression Analysis, The University of Sussex
Trostel, P., Walker, I., & Woolley, P. (2002). Estimates of the economic return to schooling for 28 countries. Labour economics, 9(1), 1-16
Walker, I., & Zhu, Y. (2001). The returns to education: Evidence from the Labour Force Surveys. Available at http://dera.ioe.ac.uk/4656/1/RR313.pdf (Accessed 09 April 2014)
Wooldridge, J.M. (2009) Introductory Econometrics: A Modern Approach. 4th edn. Canada: South-Western, a part of Cengage Learning. 85491
35
Data Reference
University of Manchester. Cathie Marsh Centre for Census and Survey Research. ESDS Government, Labour Force Survey, 2002: Teaching Dataset [computer file]. 2nd Edition. Office for National Statistics. Social Survey Division, Northern Ireland Statistics and Research Agency. Central Survey Unit, [original data producer(s)]. Colchester, Essex: UK Data Archive [distributor], October 2004. SN: 4736.
No comments:
Post a Comment