Solutions to Tutorial 5

Exercise 1
1. The results can be found below. Some results corroborate with our expectations and some do not.
Specifically, the estimate of β2 appears to be positive and highly statistically significant, as expected,
because more chronic conditions would reflect poorer health, which on average would increase medical
expenditure individuals. The age coefficients are also statistically signicant, and they are consistent
with the intuition that an additional year of age increases medical expenditure but at a declining rate,
suggesting an inverse U-shape curve, in terms of the impact of age on medical expenditure. However,
given that older people might have more serious medical conditions as they grow old and therefore
spend much more on medical products, it is not clear that one would necessarily expect the negative
coefficient on age2. The income coefficient is positive and statistically significant at the 5% level. This
is an intuitive result. The gender dummy coefficient is not statistically significant, suggesting that males
and females spend about the same on medical services, which is expected. What is most striking is
the estimate of β1 , which is counter-intuitive. This is because one would expect that having private
insurance would lowers medical expenditures because these are partly covered by the insurance itself.

2. One might expect residual heteroskedasticity, and White tests for this deliver p-values of 0, (with and
without the cross terms). However, since this will not bias the estimated coefficients, we will not pursue
the heteroskedasticity issue further here. The main issue is the coefficient of private insurancei . It
seems that there is an endogeneity problem, as one could argue that private insurance may be a choice
variable; that is, those individuals who expect to incur higher out-of-pocket medical expenditure (due
to a poor health condition) are more likely to purchase private insurance in the first place. In this case
we may write E (ui |private insurancei ) 6= 0, i.e. private insurancei is endogenous. As a result, the OLS
estimator for β1 is biased and inconsistent.
3. The variable ssiratio needs to fit the three criteria for a good instrument.
Firstly, it does not belong in the structural equation (i.e. it is uncorrelated with medical expenses).
This makes sense as we would not expect medical expenses to depend on where a person gets their
income from.
Second, it should be partially correlated with the endogenous variable private insurance. This makes
sense as those with a higher ssiratio are less likely to have private health insurance. We would expect
a negative correlation here.
Finally, we must argue that ssiratio is uncorrelated with the error term. This may not hold. It could be
argued that the higher the ssiratio, the less choice you have in getting private insurance. However, no
instrument is exactly uncorrelated with the error, so we can think about this relatively. It’s probably
the case that ssiratio is less correlated with the error than private insurance - so it is worth trying it as
an instrument.
The results below corroborate with our original expectations in that the coefficient of private insurance
becomes negative and it is statistically significant. Unlike the OLS results, this matches our expectations,
and all other estimates have the same signs and levels of significance as before.

1
4. The reasons why low income might be a good instrument for private health insurance are similar to
those given for part 3. We do not expect low income to affect medical expenses (so exclusion applies),
but we would expect it to be correlated with private insurance (so it is relevant). Further, as argued in
part 3 it is likely that low income is is less correlated with the error than with private insurance. One
advantage of including another IV is that since the number of instruments now exceeds the number
of endogenous variables, we can now test for the validity of our instruments. The IV estimates based
on two instrumental variables are reported below, and they are very similar (with respect to signs and
levels of significance) to those reported in part 3 above. The main difference is that since we have added
an additional instrument, the resulting IV estimator is more efficient and so the standard error of the
estimated coefficient of private insurancei is smaller, and similarly for other coefficient estimates.

5. We are testing the null hypothesis of exogeneity (Cov(private insurance, e)=0) vs the alternative that
this covariance is non-zero. To do this test we take the residual from the first stage regression, add
it to the structural equation and then perform a test for individual significance on this residual. The
relevant output is given below, with the first stage regression displayed first, and then the test regression
equation.  
The coefficient of the first stage residual appears to be statistically significant: δb = 1.01, s.e. δb = .273.
So tcalc = 3.705 compared to 1.96 from the Normal distribution. Therefore, the private insurance
variable appears to be endogenous, i.e. the null hypothesis of exogeneity is rejected and we conclude
that the OLS estimator is not consistent.

2
6. We test the restriction that the coefficients of ssiratioi and lowincomei equal zero in the first-stage
regression. The unrestricted model is given above in part 5 and restricted version is given below.
We wish to reject the null of weak instruments using the rule of thumb that the resulting F-statistic
needs to be larger than 10. The F-statistic from the first-stage regression equals 47.78. In particular,
the formula is given by
(SSRr − SSRu ) /h (1093.5 − 1072.2) /2
F = = = 47.78.
SSRu / (N − K) 1072.2/4820
Since the estimated statistic exceeds the value of 10, the null hypothesis of weak instruments is rejected.

7. We test the validity of the instruments by estimating the structural equation by 2SLS, saving the
residuals, and then running an auxiliary regression of these residuals on all of the exogenous variables.
The structural equation is given in part 4 and the auxilliary regression is given below.
We test the null hypothesis that the instruments are valid, vs the alternative that they are not valid.
Under the null hypothesis N × R2 from the auxilliary regression follows a χ2(1) distribution. The 5%
critical value of the χ2(1) equals 3.841, so we will reject the null if our test statistic exceed 3.841.

3
 We have N × R2 = 0.000662 × 4828 = 3.196, which is not sufficiently large to reject the null, so we
conclude that our IV estimator is consistent.

Exercise 2 Consider the following demand and supply model:

Demand: Qt = α0 + α1 Pt + α2 It + u1t , α1 < 0, α2 > 0

Supply: Qt = β0 + β1 Pt + β2 Pt−1 + u2t , β1 > 0, β2 > 0
Qdt = Qst

1. Derive the reduced form equations.

Answer:
Under the equilibrium condition, we have the following reduced form equation for Pt , which shows the
equilibrium price:

α0 + α1 Pt + α2 It + u1t = β0 + β1 Pt + β2 Pt−1 + u2t ⇒

Pt = Π0 + Π1 It + Π2 Pt−1 + v1t where

β0 − α0 α2 β2 u2t − u1t
Π0 = , Π1 = − , Π2 = , v1t =
α1 − β1 α1 − β1 α1 − β1 α1 − β1
Substituting the equilibrium price into the demand (or supply) equation yields

Qt = Π3 + Π4 It + Π5 Pt−1 + v2t where

α1 β0 − α0 β1 α2 β1 α1 β2 α1 u2t − β1 u1t
Π3 = , Π4 = − , Π5 = , v2t =
α1 − β1 α1 − β1 α1 − β1 α1 − β1
2. Deduce the structural form parameters from the reduced form equations.

Answer:

β0 = Π3 − β1 Π0 α0 = Π3 − α1 Π0
β1 = Π4
Π1 α1 = Π 5
Π2
β2 = Π5 − ΠΠ2 Π1 4 α2 = Π4 − ΠΠ1 Π2 5

3. Is the supply function identified? The demand function? What about the system as a whole?

Answer:
Since we verified before that we can obtain a unique set of estimates for all the structural form coefficients,
the system as a whole is exactly identified, which implies that every equation is exactly identified.

4
Exercise 3 One way of writing a Cobb-Douglas production function and its first-order conditions for profit
maximization is:

ln(Yt ) = β1 + β2 ln(Kt ) + β3 ln(Lt ) + e1t (1)

ln(Kt ) = α2 + ln(Pt ) + ln(Yt ) − ln(Rt ) + e2t (2)
ln(Lt ) = α3 + ln(Pt ) + ln(Yt ) − ln(Wt ) + e3t (3)
Equation (1) above defines the production function. Equations (2) and (3) come from the assumption
that each firm is maximizing profits. At period t the firm faces costs Pt (price), Rt (interest rate) and Wt
(wages). Given these costs, we assume that the firm chooses input levels Lt (Labour) and Kt (capital) and
output Yt so as to maximize profits. The existence of the errors means the firms are maximizing profits “on
average”, rather than “exactly”.

Given that Yt , Kt and Lt are simultaneously determined by equations (1), (2) and (3), and Pt , Rt and
Wt are determined from outside, it seems appropriate to treat the three equations as a simultaneous system
where Yt , Kt and Lt are the endogenous variables and Pt , Rt and Wt are the exogenous variables.

(a) Determine whether the production function (1), the capital function (2) and the labour function
(3) are identified or not.

Answer
The system has M = 3 equations and 3 endogenous variables.

The number of exogenous variables excluded from the equation must not be smaller that the number of
endogenous variables included in that equation minus one.
The number of exogenous variables omitted from the production function is 3 (Pt , Rt , Wt ). The number
of endogenous variables included in the production function is 3 (Yt , Kt , Lt ). Thus, the production function
is identified.
The number of exogenous variables omitted from the capital function is 1 (Wt ). The number of endogenous
variables included in the capital function is 2 (Yt and Kt ). Thus, the capital function is identified.
The number of exogenous variables omitted from the labour function is 1 (Rt ). The number of endogenous
variables included in the labour function is 2 (Yt and Lt ). Thus, the labour function is identified.

(b) Show that the reduced form equation for Yt (known in economics as a supply function) can be
written as    
Pt Pt
ln(Yt ) = π1 + π2 ln + π3 ln + υt ,
Rt Wt
where
β1 + β2 α2 + β3 α3
π1 = ;
1 − β2 − β3
β2
π2 = ;
1 − β2 − β3
β3
π3 = ;
1 − β2 − β3
β2 e2t + β3 e3t + e1t
vt = ;
1 − β2 − β3

Answer

Substituting equations (2) and (3) into equation (1) yields

ln(Yt ) = β1 + β2 [α2 + ln(Pt /Rt ) + ln(Yt ) + e2t ] + β3 [α3 + ln(Pt /Wt ) + ln(Yt ) + e3t ] + e1t

(1 − β2 − β3 )ln(Yt ) = β1 + β2 α2 + β3 α3 + β2 ln(Pt /Rt ) + β3 (Pt /Wt ) + β2 e2t + β3 e3t + e1t

5
    
β1 + β2 α2 + β3 α3 β2 Pt β3 Pt β2 e2t + β3 e3t + e1t
ln(Yt ) = + ln + ln +
1 − β2 − β3 1 − β2 − β3 Rt 1 − β2 − β3 Wt 1 − β2 − β3
   
Pt Pt
ln(Yt ) = π1 + π2 ln + π3 ln + υt
Rt Wt
as was required.
c) Use the result from (b) above to express β2 and β3 in terms of π2 and π3 .

Answer

Note that
1 − β2 − β3 + β2 + β3 1
1 + π2 + π3 = =
1 − β2 − β3 1 − β2 − β3
Thus,
π2
β2 = π2 (1 − β2 − β3 ) =
1 + π2 + π3
π3
β3 = π3 (1 − β2 − β3 ) =
1 + π2 + π3
(d) Explain how to use the 2SLS method to estimate the parameters.

Answer

The 2SLS method: Estimate the reduced forms of Kt and Lt :

ln(Kt ) = λ20 + λ21 ln(Pt ) + λ22 ln(Rt ) + λ23 ln(Wt ) + υ2t

ln(Lt ) = λ30 + λ31 ln(Pt ) + λ32 ln(Rt ) + λ33 ln(Wt ) + υ3t

Substitute ln(K̂t ) and ln(L̂t ) into the production function (4) and run the regression.