5.4 Empirical evidence
The neoclassical approach to discrimination produces a number of different explanations for why discrimination may exist in the labour market. Empirical analysis has tended, however, to focus not so much on testing these explanations but rather on establishing how much of an observed earnings differential between, say, men and women, can be accounted for by differences in their relative skill and education levels, their different work histories and differences in hours of work. That part of the actual differential that remains after allowing for these factors is usually attributed to labour market discrimination – it is the difference in earnings that cannot be explained by the productivity-related characteristics of the two groups. This decomposition is achieved using multiple regression analysis. The basic approach is as follows. Suppose that the earnings of men and women be determined by the following equations:
WM = a0 + a1EM
WF = b0 + b1EF
where W represents average earnings and E represents an education variable which determines earnings. In the first equation W M represents the earnings of men, while E M represents, say, the average number of ‘A’ levels held by men. In the second equation W F the average earnings of women, while E f is the average number of ‘A’ levels held by women. The terms a 0 and b 0 are the intercepts of the two equations. The observed average gender earnings differential – the difference in earnings between the two groups – can then be broken down by subtracting one equation from the other:
(WM − WF) = (a0 − b0) + a1EM − b1EF + (b1EM − b1EM)
The term (b1EM − b1EM) has been artificially added to this equation. The equation can, as a result, be re-arranged to give us the expression:
(WM − WF) = (a0 − b0) + b1(EM − EF) + (a1 − b1) EM
The equation shows that the average wage differential – the difference in average earnings between men and women – is made up of three components.
The first term:
(a0 − b0)
is the part of the gender differential that is due to differences in earnings that take place on entry to the labour market. This will reflect differences in pre-entry human capital investments and/or pre-entry discrimination.
The second term:
b1(EM − EF)
shows the contribution that differences in productivity-related characteristics between men and women make to the gender earnings differential. In other words, it shows the effect that gender differences in the average level of education have on the average earnings differential. This component represents that part of the earnings differential that does not reflect discrimination.
The final term:
(a1 − b1)EM
measures the impact that unequal treatment has on the average wage differential. This is measured by the difference between the coefficients in the two equations on the variable measuring education. These coefficients show how education affects an individual's earnings. This term measures what happens when the labour market rewards productivity-related characteristics in different ways. It is this term, therefore, that is used as a measure of discrimination – women's productivity-related characteristics are treated differently from those of men for reasons that have nothing to do with the characteristics themselves. So, for example, suppose that women with four ‘A’ levels earn less than men with four ‘A’ levels (assuming all other things equal). This approach says that this can only come about because of discrimination. The extent of this discrimination is captured by the amount that the woman is paid less than her male comparator, which, in a model like the one above, is reflected in a smaller coefficient on the education variable in the female earnings equation compared with the earnings equation estimated for men.
The following example provides a demonstration of how the model can be estimated. Wright and Ermisch (1991) conducted an analysis of discrimination against women in the UK using data from the 1980 Women and Employment Survey. A total of 2094 employed women aged between 16 and 59 were interviewed in a nationally representative sample. Alongside this survey of women, husbands of those 1868 women who were married at the time were also interviewed.
Everyone was asked, among other things, their wage per hour, their place of residence, and any educational qualifications they had. Using a statistical procedure developed in a previous study (Wright and Ermisch, 1990), a variable for the potential experience of individuals was calculated. Since, unfortunately, information on the work experience of individuals was not collected in the survey, this was estimated using other information. Variables for education were specified to model the impact on wages of individuals holding CSEs, ‘O’ levels, ‘A’ levels, degrees or any other qualifications. The wage equations for males and females are reported in Table 5. Note that for ease of exposition, a number of the additional variables used in the study, such as those representing the regions in which individuals reside, are not reported here.
|(sample size 1868)
|(sample size 2094)
On examining the two wage equations we need to check a number of features. First, we need to check the sign of the coefficients for each variable. In the wage equations for both males and females all the variables shown have a positive sign. This means that all these variables have a positive influence on wages. For example, the positive sign for potential experience means that potential experience is positively correlated with wages – there is a positive premium on experience. As potential experience increases, then, on average, both males and females should get a higher wage. The coefficients for all of the education variables are also positive for both genders. For example, the coefficient on the variable for males with ‘O’ levels is 0.203; this means that there is a premium in terms of higher wages for male workers obtaining ‘O’ levels compared to someone with no qualifications. The comparable value for females is 0.116.
Second, we need to check the statistical significance of the coefficients. To do this the t-statistics associated with each coefficient can be examined. A t-statistic greater than 1.96 for a particular coefficient means that we can have 95 per cent confidence in the statistical importance of that coefficient. In Table 5, with the exception of a t-statistic of 1.60 for ‘Other’ qualifications, all the reported t-statistics for males are higher than 1.96. A t-statistic of 0.58 for ‘Other’ qualifications and a t-statistic of 1.52 for potential experience are reported for females. Econometricians often estimate equations in which some coefficients have low t-statistics, but this reduces their confidence in the results obtained. We might also be somewhat cautious about the two wage equations in Table 5 since the R2's are 0.249 and 0.224 for males and females respectively.
This means that over 70 per cent of the variation in wages is not explained by these equations.
The key thing to note about these equations is that, in all cases, the coefficients for males are larger than the coefficients for females. Take, for example, the wage premium for ‘A’ levels. In the equation for males the premium on ‘A’ levels (the size of the coefficient) is 0.316, but in the equation for females it is only 0.210. The premium on potential experience shows an even wider disparity with a coefficient value of 0.246 for males and only 0.044 for females.
A formal measure of these differences in the size of coefficients is calculated by using the method of decomposition explained earlier in this section. Look back and make sure you have grasped that the differential between male and female earnings can be explained in terms of two main components:
the part of the earnings differential which is due to differences in productivity
the part of the earnings differential which is due to male and female workers of equal productivity being rewarded differently.
For the wage equations reported in Table 5 and the additional variables not reported in the table, Wright and Ermisch derive the estimate that 11.8 per cent of the wage differential is attributed to (1) and 88.2 per cent to (2). This means that unequal rewards for the same productivity are found to be more important when explaining wage differentials than differences in productivity. It should be noted, of course, that these results are qualified by the small R2's of their wage equations. It should also be noted that Wright and Ermisch estimate a number of other wage equations with different specifications, which give less weight to the importance of discrimination. Nevertheless, even taking these cautious notes into account, the results reported in Table 5 provide a revealing insight into the extent to which discrimination takes place against women in the labour market.
It is more difficult to obtain empirical evidence about racial discrimination than for sex discrimination. Substantial earnings differentials have been found to exist between white and ethnic minority workers, with the latter earning about 10 per cent less than whites (Blackaby et al, 1994). A significant part of this differential reflects the occupational and industrial segregation of minority ethnic groups. Indeed, while there is evidence that minority ethnic groups face wage discrimination and, in particular, face lower rates of return to education and general training, it is discrimination in terms of occupational access which is found to be of most significance (McNabb and Psacharopoulos, 1981).
However, this type of analysis only compares the earnings of people in work and minority ethnic groups are significantly more likely to be unemployed than white workers. This raises two issues.
Why do workers from minority ethnic groups face worse employment prospects than white workers?
What impact does this have on the earnings differential between the two groups?
A study of unemployment among Britain's minority ethnic groups (Blackaby et al, 1995) found that although employees from minority groups had less favourable characteristics (in terms of the attributes that affect the likelihood of employment, such as age, education and so on), the main reason for their different unemployment experiences was discrimination. However, significant differences were found between the ethnic groups considered. For example, relatively high unemployment among workers of West Indian origin reflected their unfavourable characteristics rather than discrimination. In contrast, unemployment among workers of Indian descent was primarily the result of discrimination. They experienced relatively more unemployment even though their characteristics were, in fact, more favourable in terms of the likelihood of finding employment than those of white workers. Finally, the analysis indicated that discrimination against workers of Pakistani and Bangladeshi origin was greater than against those of Indian descent. The authors attribute the latter finding to the fact that workers of Pakistani and Bangladeshi descent, ‘have reacted to discrimination in a different way by choosing to be more isolated and have adopted an economic structure which is more autarchic compared to other groups. Greater economic disadvantage is the consequence of this’ (Blackaby et al, 1995, p. 25).
The same authors also examine the interaction between unemployment and wage discrimination among ethnic minority workers. They find that the limited employment prospects faced by ethnic minority workers are significantly more severe than the earnings disadvantage they face. As unemployment in Britain increased in the 1980s and early 1990s, minority ethnic groups suffered disproportionately. Even ethnic minority employees with favourable productivity-related characteristics (those with better earnings potential than white workers) have become unemployed, thereby increasing the wage gap between white and non-white employees.