Resultados de la Comparación de Entrevistas Estructuradas

3.4.1

E stim atin g th e R etu rn s to E du cation

There are a number of approaches to estim ating the returns to education as we saw in Chapter 2. Following the approach outlined in th at chapter we begin by using a two equation system

Wi = SiPi + ATi/?2 4- Ui (3.1)

Si = Z'i'y + Vi (3.2)

where Si is years of schooling (full-time years of education), Wi is the log of the real hourly wage rate, X i and Zi are vectors of exogenous observed individual characteristics, /?i is the return to education, and Ui and Vi are a pair of residuals. OLS estim ation of equation (3.1) gives rise to an unbiased estim ate of the return to education if Ui and Vi are uncorrelated, th a t is if Si

is exogenous in the wage equation (E{siUi) = 0).

Instrum ental variable approaches identify a set of exogenous variables th a t affect the education decision, but not earnings controlling for educa­ tion. In order for the model to be identified we need variables in our Zi

can be legitimately be left out of the earnings equation (3.1) while being significant determ inants of our education variable in equation (3.2). Thus we have Z'i = (X ', W/) where Wi is a vector containing at least one variable for identification. In our study we use family background variables, family com­ position variables and variables identifying the type of school the individual last attended. Our schooling equations also include year of birth dummies

to capture cohort effects. Our exogenous explanatory variable in the wage equation (X^), consist of gender, race variables for Australian born individu­ als, dummy variables identifying English proficiency for individuals for whom English was a second language (ESL) as well as regional and year dummy variables^. Experience variables are also included in our wage equation, but we treat these as endogenous^. We also include the a variable controlling for the number of siblings the individual has. We initially carry out IV estim a­ tion treating our years of education variable as a continuous variable. This is equivalent to estim ating the following wage equation

Wi = SiPi 4- 4- T]i (3.3)

where are the residuals from OLS estimation of equation (3.2), and E{sirji) =

0 by construction. A Hausman t test of the exogeneity of schooling is given by testing û; = 0 in equation (3.3)®.

Previous studies looking at the returns to education in Australia suggest th a t education is not a continuous variable^. We have two measures of educa­ tional outcomes both of which are ordered. Hence for both of these different measures we also use a latent variable model of the form

~ ^i'y d* (3-4)

where

Sij = 1 if fJ'j—i ^ fij (3.5)

where j = 0 ,1,2,3... and j i j - i < [ i j . The education equation is now estim ated

as an ordered probit and the param eter estimates are used to calculate the

®In earlier versions of this work, we also interacted schooling with a tim e trend to see how th e returns to education had changed over tim e, controlling for year effects. There was no evidence o f changes in the returns to education for women, and only marginal evidence of a decline for men over the period 1985 to 1991.

^Experience in our data has been defined as age minus years of education minus five. Since years of schooling is endogenous in our set up, years of experience is also endogenous.

®See Sm ith and Blundell [148].

usual Heckman [97] selection adjustm ent term

t -

z'â)

m

• $(% + ! - Z'a) - - Z'a) ^ ■ •'

where the /i^’s and 7 are the estimates obtained from the ordered probit maximum likelihood procedure, and <^(.) and ^ (.) are the normal probability distribution and normal cumulative distribution functions respectively. We can then estim ate the following wage equation

(3.7) where Si is now either years of education or a vector of dummy variables identifying the person’s highest qualification. This specification my be less robust th an equation (3.3) because we require an additional assumption of normality. In estim ating both equations (3.3) and (3.7) our standard errors are corrected to take account of the generated regressor in the equations as well as heteroscedasticity^®.

3.4.2

E du cation and G ender W age D ifferentials

The mean difference in the observed wages of men and women in terms of log differences, or gender gap (g) is given by

2

3^0

where Xm and Xf are vectors containing the means of all the explanatory vari­ ables in our male and female wage equations (except selection terms which have a mean of zero for the whole sample) and (5m and (3f are the correspond­ ing estim ated coefficient vectors. Following the approach of Stewart [154]^^ and Juhn, M urphy and Pierce [107] we can rewrite this expression as

g = ( 3 m - + x'fiPm ~ M = 9c + 9p (3.9)

^°See Arellano and Meghir [4] for details.

which decomposes the observed gender wage differentials into two effects^^. The first is the difference in observed wages which arises because men and women have different observed characteristics, for instance education and labour m arket experience. The second is the differences in observed wages which is a result of men and women being “paid” differently for a given set of characteristics. This is the estimated differential which exists once back­ ground has been controlled for or ceteris paribus gender wage differential. If observed gender wage differentials primarily reflect differences in observed characteristics then the policy response will be different than if they prim ar­ ily reflect differences in the “price” paid for the observed characteristics of women. The mean gender wage gap of any subgroup s, of our sample, for instance individuals with a particular educational qualification, can be calcu­ lated by replacing the mean characteristics of males and females with those of the subgroup s of interest, and x'^f in equation (3.9).