Estimation of Educational Borrowing Constraints Using Returns to ...
Estimation of Educational Borrowing Constraints Using Returns to ...
Estimation of Educational Borrowing Constraints Using Returns to ...
- No tags were found...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
educational borrowing constraints 167<br />
and<br />
′ ′ ′<br />
V1i p a1 log {exp [ log (R i) XWibW XL1ibL1 v 1i] X Cib C}<br />
′<br />
a 2 a 3log (R i) XTibT1 n 1i, (25)<br />
where a1, a2, and a3<br />
are scalars (as defined in [23]) that do not vary<br />
across individuals.<br />
Though (24) and (25) are complicated, these expressions are close<br />
<strong>to</strong> a familiar linear index model. A nonlinearity arises inside the logarithm<br />
in the first term <strong>of</strong> (25) and captures the interaction between the<br />
′<br />
borrowing rate R i and schooling costs X CibC. No such interaction exists<br />
between interest rates and forgone earnings, which are represented by<br />
′ ′<br />
X ˜<br />
WibW XL0ibL0 v0i. This feature <strong>of</strong> the model delivers identification<br />
<strong>of</strong> the parameters <strong>of</strong> interest.<br />
We estimate two versions <strong>of</strong> the model. The first restricts variation in<br />
R i <strong>to</strong> be determined through particular sets <strong>of</strong> observables, such as race<br />
or family income. The second version treats R i as a variable known <strong>to</strong><br />
the individual but not observed by the econometrician. Identification<br />
<strong>of</strong> each case is discussed separately.<br />
˜<br />
1. Case I: R i Determined by Observed Characteristics<br />
The first approach assumes the absence <strong>of</strong> persistent, unobserved individual<br />
influences on wages ( vSi<br />
p 0) and assumes that R i varies only<br />
with observables. The empirical content <strong>of</strong> this version <strong>of</strong> the model is<br />
closely related <strong>to</strong> the regression approach presented in Section VF. To<br />
see why, suppose that local labor market and wage variables are not<br />
′ ′<br />
determinants <strong>of</strong> schooling value ( X ˜<br />
WibW XLSibSL<br />
p 0 for both S p 0<br />
′<br />
and S p 1). Let R i be parameterized by the index XRibR, and assume<br />
′<br />
that the schooling costs index, X CibC, varies only with the presence <strong>of</strong><br />
a local college. Equation (25) shows that log (R i ) enters the value func-<br />
′<br />
tion through the terms log {exp [ log (R i)] X Cib C} and a 3 log (R i)<br />
.<br />
Since the observable variables that determine R i are also included in<br />
the set <strong>of</strong> observables governing tastes, X Ti , and since they are all dummy<br />
variables (such as race), a 3 log (R i)<br />
is not separately identified from<br />
′<br />
XTibT1. Therefore, identification <strong>of</strong> log (R i)<br />
comes completely from the<br />
′<br />
term log {exp [ log (R i)] X Cib C}<br />
. In addition, since X Ci contains only<br />
a constant and an indica<strong>to</strong>r variable, testing for interest rate heterogeneity<br />
(i.e., bR<br />
p 0) is identical <strong>to</strong> testing for interactions between the<br />
presence <strong>of</strong> a local college and the variables determining the borrowing<br />
rate (such as race) and almost identical <strong>to</strong> tests for the ad hoc interactions<br />
estimated in the schooling regressions <strong>of</strong> table 7. The main<br />
difference is that the schooling outcome is modeled here as a discrete<br />
variable.