FECO - Exercise session IV

Question 1

Recall the regression model from FECO Ex3. The OLS estimates for a sample of size n = 500
are as follows:

ln(s\
salaryi ) = 3.63 + 0.012 ln(assessment testi ) − 0.035 grade metricsi
(0.92) (0.04) (0.005)
+ 0.21 internshipi + 0.46 exchangei + 0.08 internshipi × exchangei ,
(0.06) (0.23) (0.23)

where standard errors are given in parentheses. Given the sample size, you may use quantiles
from the standard normal distribution as critical values:
α zα (α-level quantile) α zα (α-level quantile)
0.005 −2.58 0.995 2.58
0.025 −1.96 0.975 1.96
0.050 −1.64 0.950 1.64
0.100 −1.28 0.900 1.28

Note that since the standard normal distribution is symmetric, the corresponding quantiles for
the lower tail critical value can always be constructed multiplying the value by −1 (e.g., the
value of the 0.025 quantile is −1.96). As in the corresponding exercise of FECO Ex3, assume
again the error term is statistically independent of all covariates.

(a) Test whether ln(assessment testi ) and grade metricsi are significant at the 1% significance
level.

(b) Test whether the (average) salary elasticity w.r.t. the assessment test is 0.5% (vs. that
it is not) at the 5% significance level.

(c) Suppose you define a new variable noexchangei as negation of exchangei , i.e. noexchangei =
1 − exchangei , and repeat the above regression using noexchangei instead of exchangei
(the rest remains identical to before). Does the estimate of β4 change? What about the
estimate of β5 and the standard errors of both estimated coefficients?

(d) Suppose you hold an (economic) prior belief that exchangei can only be non-negatively
associated with ln(s salaryi ). (i) What kind of H0 and H1 would you use to test this prior
belief? (ii) Carry out the corresponding t-test at the 5% significance level. (iii) How could
you modify H0 and H1 if you are looking for evidence of a positive linear association? (iv)
Carry out the corresponding t-test at the 5% significance level.

1
Question 2

Suppose you estimate the following linear model:

ln(wagei ) = β0 + β1 jci + β2 univi + β3 experi + ei

to study the difference in returns to junior college (jci lasting for a maximum of 2 years) and
four-year college (univi ), while experi measures months in workforce and wagei monthly wages
in dollars. The OLS estimates for a sample of n = 6, 763 observations are:

\ ) = 1.472 + 0.0667 jc + 0.0769 univi + 0.0049 exper + ei

ln(wagei i i
(0.021) (0.0068) (0.0023) (0.0002)

Given the sample size, you may use again the critical values from exercise (i).

(a) Construct a 95% Confidence Interval for the coefficients β1 and β2 .

(b) Test whether the returns of junior college are at least as high as the ones from higher
university, i.e. H0 : β1 ≥ β2 vs. H1 : β1 < β2 . The estimated variance covariance matrix
of the estimated coefficients is given:

jc univ exper _cons

jc .00004663
univ 1.928e-06 5.330e-06
exper -1.718e-08 3.933e-08 2.480e-08
_cons -.00001741 -.00001573 -3.105e-06 .00044353

(c) Can you think of a way to modify the linear regression model in (a) to test the same
hypothesis using only a t-test with one estimated coefficient and its corresponding stan-
dard error (assuming that people may in fact attend junior college or university alone, or
both)?

Question 3

Using the Californian test score data CASCHOOL.dta, you are interested in analyzing the re-
lationship between districts’ test scores and their average class sizes (measured as average
student-teacher ratio), controlling also for avginci , the average houshold income in district i,
and meal pcti , the average share of students who receive free lunch at school. That is, estimate
the following model:

testscri = β0 + β1 stri + β2 str2i + β3 str3i + β4 avginci + β5 meal pcti + ei

(a) (i) How could you test whether it is important to allow for a nonlinear profile in stri

2
 in this context? State the null and the alternative hypothesis using the matrix-vector
notation from the lecture, i.e. as Rβ = c defining what these quantities are in this case.
(ii) Carry out the test at the 5% significance level (Hint: you can do this in Stata by
using the post estimation command test after the execution of the regression, type help
test into the command window for more information).

(b) The data also contains the variable el pcti , which indicates the share of English learners
among the students in the district. Generate a variable HiEL = 1 for districts with more
than > 10% English learners, and zero otherwise. A researcher suggests that districts
with a high share (larger than 10%) of students whose first language is not English might
exhibit a different average test score and class size relationship compared to (a). (i)
Suggest an alternative specification to the one in (a) to test this claim empirically. (ii)
State the corresponding hypothesis for this claim. Use again the matrix-vector notation
Rβ = c to state your hypothesis, defining what these quantities are. (iii) Carry out the
test at the 5% significance level and state your conclusion. What additional information
does the marginal significance level (p-value) convey?
Remark: the output from the OLS regressions to answer the above questions are as follows
(strsq and strcu are the quadratic and cubic transformation of str):
. reg testscr str strsq strcu avginc meal_pct, robust

Linear regression Number of obs = 420

F(5, 414) = 327.45
Prob > F = 0.0000
R-squared = 0.7925
Root MSE = 8.7308

------------------------------------------------------------------------------
| Robust
testscr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
str | 70.3433 24.41211 2.88 0.004 22.35616 118.3304
strsq | -3.734898 1.232892 -3.03 0.003 -6.158406 -1.31139
strcu | .0645509 .0205632 3.14 0.002 .0241296 .1049723
avginc | .5181386 .0801448 6.47 0.000 .3605972 .6756801
meal_pct | -.5071565 .0217612 -23.31 0.000 -.5499328 -.4643803
_cons | 238.758 159.6049 1.50 0.135 -74.97903 552.495
------------------------------------------------------------------------------

. test strsq strcu

( 1) strsq = 0
( 2) strcu = 0

F( 2, 414) = 6.76

3
 Prob > F = 0.0013

. gen HiEL=0

. replace HiEL=1 if el_pct>=10

(192 real changes made)

.
. gen strEL = str*HiEL

. gen strsqEL = strsq*HiEL

. gen strcuEL = strcu*HiEL

. reg testscr str strsq strcu strEL strsqEL strcuEL avginc meal_pct, robust

Linear regression Number of obs = 420

F(8, 411) = 226.12
Prob > F = 0.0000
R-squared = 0.8065
Root MSE = 8.4625

------------------------------------------------------------------------------
| Robust
testscr | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
str | 51.39552 25.3096 2.03 0.043 1.643104 101.1479
strsq | -2.687655 1.286995 -2.09 0.037 -5.217568 -.1577424
strcu | .0459194 .0215787 2.13 0.034 .003501 .0883379
strEL | 2.582287 3.278859 0.79 0.431 -3.863137 9.027712
strsqEL | -.2775349 .3191828 -0.87 0.385 -.9049693 .3498995
strcuEL | .0066675 .007738 0.86 0.389 -.0085435 .0218784
avginc | .6402001 .0833372 7.68 0.000 .4763797 .8040205
meal_pct | -.4324647 .0258129 -16.75 0.000 -.4832064 -.3817229
_cons | 345.63 163.997 2.11 0.036 23.25235 668.0076
------------------------------------------------------------------------------

. test strEL strsqEL strcuEL

( 1) strEL = 0
( 2) strsqEL = 0
( 3) strcuEL = 0

F( 3, 411) = 10.26
Prob > F = 0.0000