Economics Honors Exam 2010 Solutions: Econometrics, Question 9

We have a random sample of size n from some population. For each individual i,
there are variables (Yi , Zi , Ai ), where we interpret Zi as a noisy measure of Ai . We
observe (Yi , Zi ) for i = 1, . . . , n, but we do not observe Ai . The structural model is
Yi = 0 + 1 Ai + Ui
Zi = Ai + Vi
(i = 1, . . . , n),
where the latent (unobserved) random variables Ai , Ui , and Vi have mean 0 and
are all uncorrelated with each other. In addition, we assume that 1 > 0.
a. (10 points) Work out the covariance matrix for (Yi , Zi ) (i.e., express the two
variances and one covariance in terms of the model parameters 1 , Var(Ai ), Var(Ui ),
Var(Vi )).
Answer:


 

Var(Zi )
Cov(Zi , Yi )
Var(Ai ) + Var(Vi )
1 Var(Ai )
=
Cov(Zi , Yi )
Var(Yi )
1 Var(Ai )
12 Var(Ai ) + Var(Ui )

b. (25 points) Consider the linear predictors

E (Yi | 1, Zi ) = 0 + 1 Zi ,
E (Zi | 1, Yi ) = 0 + 1 Yi .
Show that
1 1 1/1 .
Provide a procedure that uses the data on (Yi , Zi ) for i = 1, . . . , n to provide estimates
of these lower and upper bounds for 1 . Explain why the estimates are consistent.
Answer:
1 =

Var(Ai )
Cov(Zi , Yi )
= 1
1
Var(Zi )
Var(Ai ) + Var(Vi )

Var(Yi )
12 Var(Ai ) + Var(Ui )
Var(Ui )
1/1 =
=
= 1 +
1
Cov(Zi , Yi )
1 Var(Ai )
1 Var(Ai )
1

Use the least-squares estimates:

1 =
1/
=

sample Cov(Z, Y )
,
sample Var(Z)
sample Var(Y )
,
sample Cov(Z, Y )

where, for example,

n

1X
1X
1X
sample Cov(Z, Y ) =
Z i Yi
Zi
Yi .
n i=1
n i=1
n i=1
These estimates are consistent because, under random sampling, sample means converge to population means, and so sample covariances converge to population covariances.
c. (25 points) Now suppose that there are two noisy measures of Ai , with
Yi = 0 + 1 Ai + Ui
Zi1 = Ai + Vi1
Zi2 = Ai + Vi2
(i = 1, . . . , n),
where the latent (unobserved) random variables Ai , Ui , Vi1 , and Vi2 have mean
0 and are all uncorrelated with each other. Show that now 1 is identified, and use
your analysis to provide a consistent estimator of 1 using the data on (Yi , Zi1 , Zi2 )
for i = 1, . . . , n.
Answer:
The covariance matrix for Zi1 , Zi2 , Yi is

Zi1
Var(Ai ) + Var(Vi1 )
Var(Ai )
1 Var(Ai )
.

Var(Ai ) + Var(Vi2 )
1 Var(Ai )
Cov Zi2 =
2
Yi

1 Var(Ai ) + Var(Ui )
So 1 is identified from
1 =

Cov(Zi1 , Yi )
Cov(Zi1 , Zi2 )

or 1 =

Cov(Zi2 , Yi )
.
Cov(Zi1 , Zi2 )

We can obtain a consistent estimator for 1 by using sample covariances to estimate

population covariances (as in (b)):
sample Cov(Z1 , Y )
1 =
sample Cov(Z1 , Z2 )

sample Cov(Z2 , Y )
or 10 =
.
sample Cov(Z1 , Z2 )
2