model, we see that, except for some of the nonblack min

ority variables, all of the race and sex variables were

statistically significant in all of the models in which

they were used.

34/

Furthermore, they were all signifi

This is a very strong

cant at a 1 percent alpha-level.35/

and important result. It means that when we account for the variation in salary which can be explained by variations in years of government service, age, education, geographic location, and organizational unit, a significant portion of variation remains unexplained which can be

accounted for by the race and sex variables.

That is, race

and sex are important determinants of an employee's salary in the Labor Department; they account for salary differences that are not explained by differences in age,

34/

No race and sex variables were used in the models with the finest possible race and sex breaks in themselves. They were used wherever they were necessary to give a finer break. To see in which models they appeared, refer to the table on page 88.

35/

In this case, statistical significance is affirmed if a hypothesis stating that a given variable is not significant can be rejected. The alpha-level is the probability of rejecting a hypothesis that was true.

experience, and education alone.

This conclusion is even more important because it

speaks to the basic research question which the correlation analysis was seeking to answer. That question was; having demonstrated through straight quantitative techniques that age, education, and experience considered individually do not account for the major portion of the salary gap between minority and female employees and all employees, do they, when considered simultaneously, explain the major portion of these salary differences? has been shown, the question can now be answered in the Minority and female employees in the Department

negative.

As

do earn less than all employees, and a major portion of this salary gap exists just because of their race or sex. The magnitude of the salary differences explained by race and sex variables in the regression models is shown by the table on page 97. Only statistically sig

nificant differences are shown.

For example, considering

the salary model for all employees using the race-sex dummies giving sex within race, we see that after computing a salary prediction for an employee by plugging in specific values for all of the variables not relating to race and

Magnitude, in Dollars, of Race and Sex Coefficients by Regression Model *

-4206

-1650

-2805

-3067

+

-1270

+

sex, we must make the following adjustments if the employee

is not a white male: (1) subtract $5,119 if she is a black female, (2) subtract $4,131 if she is a white female,

(3) subtract $3,729 if she is a nonblack minority female, (4) subtract $3,085 if he is a black male, and (5) subtract

$1,178 if he is a nonblack minority male.36 And these

differences are after any differences in age, education, and length of government service have been accounted for. The salary gap is thus seen to be nearly as large as when the other explanatory factors are not considered.

Finally, consider an employee who joins the Department of Labor at age 25 with a bachelor's degree. The employee is now 30 years old, possesses 5 years of government experience, and has earned graduate credits equivalent to another year of education. We can use the various regression

models to predict the employee's salary, as follows:

37/

36/

The result given for a nonblack minority male was not statistically significant at a 5 percent alpha-level.

37/ The regression models for sex within race were used to develop this comparison. They were used as outputted by the regression program; that is, before elimination of any of the variables. For purposes of the comparison, the employees were assumed to work for either ESA or LMSA in the National Office, although similar results would have been obtained from other assumptions.