KØBENHAVNS UNIVERSITET

Assignment 2

Assignment 2 - Relationship Between Residential Investments and Tobin’s Q

September 3, 2023
Contents
1 Introduction 3

2 Econometric theory 3
2.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Misspecification tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.4 Granger causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.5 Structural VAR Model and Impulse-Responses . . . . . . . . . . . . . . . . . . . . . . 5

3 Description of the data 5

4 Empirical analysis 6

5 Discussion 9

6 Conclusion 9

2
1 Introduction
In our report, we analyze the dynamic relationship between Tobin’s Q and the change in residential
investments. Tobin’s Q measures the ratio between the price of existing buildings and the construction
cost. The analysis is made since theory suggests that Tobin’s Q drives residential investments. The
answer to this question is derived from VAR models since they are useful for analyzing when the
causal relationships between variables are unknown. Moreover, we use Granger causality to detect
the nature of the causal relationship between Tobin’s Q and residential investments, and at the end
impulse responses to show the effects, on residential investments, of a change in Tobin’s Q.

In the end, we conclude that our analysis substantiates the theory that Tobin’s Q drives residential
investments.

2 Econometric theory
2.1 Model
To analyze the dynamic relationship between changes in residential investments and Tobin’s Q the
′
natural starting point is VAR models. Where we can define the model as, Zt = (z1t , z2t , ..., zpt ∈ Rp
and provides us to examine the dynamic relationship between variables. Our analysis is a bivariate
case with a 2-dimensional vector Zt = (∆4 invt , ∆4 Qt )′ and the bivariate VAR(k) model

Zt = µ + Π1 Zt−1 + Π2 Zt−2 + ... + Πk Zt−k + ϵt , t = 1, 2, ..., T, (1)

d
with ϵt | It−1 = N (0, Ω), where k represents the amounts of lags, µ is the constant factor, and ∆4"Inv#t ,
ϵ1t
∆4 Qt are the dependent variables in period t. In the bivariate case, ϵt is a 2x1 vector ϵt = ,
ϵ2t
then Ω is a 2x2 symmetric matrix and the information set is It−1 = {Zt−1 , Zt−2 , ...}. The choice
of ∆4 Invt and ∆4 Qt as dependent variables will be further elaborated. For the VAR model to be
consistent the assumptions of stationarity and weak dependence have to hold.

2.2 Estimation
To estimate a VAR(k) model we can estimate each equation by OLS. We have the OLS estimator
given by the two quantities !
T
X T
X
′
Φ̂ = Zt Z̃t−1 Z̃t−1 Z̃t−1 (2)
t=1 t=1
T
1X ′
Ω̂ = ϵ̂t ϵ̂ , where ϵ̂t = Zt − Φ̂Z̃t−1 (3)
T t=1 t

which is the same as the MLE of {Φ, Ω}. We have to be aware of inference for our MLE to be
consistent and asymptotically normally distributed hence we assume ϵt is i.i.d. (0, Ω). Furthermore,
the assumptions earlier stated of stationarity and weak dependence have to hold. For the estimator
we have to note that the error term might be heteroskedastic, in that case, we will have to use a
heteroskedastic robust covariance matrix instead.

2.3 Misspecification tests

To identify the number of lags we use the Box-Jenkins identification procedure. By doing a graphical
analysis of the ACF and PACF we get an indication of the number of preferred lags. The PACF is

3
showing significant lags, while ACF shows if there is persistence.

After the estimation of the first model, we are interested in removing insignificant lags to make
the model more parsimonious. To test for insignificant lags we use t-test and define the null- and
alternative hypotheses:
H0 : θ = 0

HA : θ ̸= 0

The t-test follows a Gaussian distribution, in fear of removing too many terms and risk getting
autocorrelation, we here use a critical value of 1.

To further determine the sufficient lag length for the data in our VAR model, we use the wald-test.
Wald-test is preferred here since it can be made robust by using the QMLE variance and is important
if we have heteroscedasticity and non-normality. The Wald tests are performed on the following null
and alternative hypotheses
H0 : θe1 = θe2 = ... = θel = 0

HA : θe1 = θe2 = ... = θel ̸= 0

with {e1 , e2 .., el } being the amount of excluded variables. The Wald test is χ2 distributed, with a
significant value of α = 0.05, the critical value depends on the degree of freedom.

We have to ensure the preferred model is stationary and weakly dependent. From theorem 6.21 it’s
known that Zt in a VAR(k) model is stationary and weakly dependent if all pk eigenvalues of the
companion matrix are inside the unit circle. The eigenvalues λ1 , λ2 , ..., λpk are given by the solutions
to
|Π̃1 − λIpk | = 0.

2.4 Granger causality

The Granger causality test examines whether the lagged values of one variable help predict another
variable in the VAR model. The test consists of a wald-test, in which we force an exclusion restriction
on the lagged variables.

D4Invt = µ1 + Πa1,1 D4Invt−1 + ... + Πak,1 D4Invt−k + Πa1,2 D4Qt−1 + ... + Πak,2 D4Qt−k + ϵ1t (4)

D4Qt = µ2 + Πb1,1 D4Invt−1 + ... + Πbk,1 D4Invt−k + Πb1,2 D4Qt−1 + ... + Πbk,2 D4Qt−k + ϵ2t (5)

Depending on the direction, we can define the hypothesis as:

H0 : D4Q ↛ D4Inv : Π121 = ... = Πk21 = 0

H0 : D4Inv ↛ D4Q : Π112 = ... = Πk12 = 0

And for the alternative hypothesis, is the same for both cases:

HA : Π1 = ...Πk ̸= 0

With a significant value of α = 0.05 and a χ2 distribution, the critical value depends on the df.
The null hypothesis claims that the lag variables are insignificant, thus predicting that there is no
causality. Vise versa the alternative hypothesis states that we cannot reject causality.
1 Heino Bohn Nielsen. A Course in Time Series Econometrics. page 139.

4
2.5 Structural VAR Model and Impulse-Responses
Assuming that the direction of the causality follows the macroeconomic theory, such that D4Q →
D4Inv, we can define the SVAR model:

D4Invt = µ1 + Πa1,1 D4Invt−1 + ... + Πak,1 D4Invt−k + Πa1,2 D4Qt−1 + ... + Πak,2 D4Qt−k + ΦD4Qt + ϵ1t
(6)
D4Qt = µ2 + Πb1,1 D4Invt−1 + ... + Πbk,1 D4Invt−k + Πb1,2 D4Qt−1 + ... + Πbk,2 D4Qt−k + ϵ2t (7)
In the SVAR model, it will be possible to examine the effect of Tobin’s Q in the same quarter. This
opens up the option for an orthogonalized impulse responses as the SVAR will be orthogonal by con-
struction2 .

Considering the SVAR model, the impulse response analysis consists of creating a shock to the ϵt ,
where can define the responses as:
∂Zt ∂Zt+1 ∂Zt+2
= Ip , = C1 , = C2 , . . .
∂ϵ′t ∂ϵ′t ∂ϵ′t

And under the assumption of a stable model, we can define the moving average representation as:

Zt = ϵt + C1 ϵt−1 + C2 ϵt−2 + . . . Ct−1 ϵ1 + C0

It is clear to see that the coefficients are the same as the impulse response.
An impulse response of the SVAR model is useful for us to illustrate the transition of the shock.

3 Description of the data

The data set is taken from the database of the Mona which is maintained by the Danish Central
bank, and contains quarterly data for residential investment (Inv ) and Tobin’s Q (Q). The data set
also provides four transformations of these variables: DInv, DQ, D4Inv and D4Q. The latter two
represent the changes in each variable compared to the previous year (4 periods), and are obtained
using the formulas:
∆4 Invt = Invt − Invt−4 (8)
∆4 Qt = Qt − Qt−4 (9)
The data set consists of values going from 1971 period 1 to 2019 period 4.

Figure 1: D4Inv Figure 2: D4Q

We can see that both of the variables are stationary. From the graph it can be seen, that ensuring
normality, might be an issue, but it would still be preferable to include a dummy of 2000(1), due to
the dotcom crash.
2 Coruse 12p.147
v

5
4 Empirical analysis
As mentioned before, we will use the transformations D4Q and D4Inv, to reach stationarity. To find
the correct amount of lags in our VAR model, we make a graphical analysis of the ACF and PACF
graphs.

Figure 3: Autocorrelation function (ACF) and partial autocorrelation function (PACF)

It is apparent that in both series there is some clear persistence from earlier periods, and consid-
ering the PACF, we can let the number of lags be 9, as we do not risk auto-correlation.
We can then define our bivariate VAR(9) model as:

Zt = µ + Π1 Zt−1 + Π2 Zt−2 + ... + Πk Zt−9 + θ1 D2000(1) + ϵt , t = 1, 2, ..., 9, (10)

As we aim for a parsimonious estimation, we want to reduce our model without risking auto-
correlation. Therefore, we create a model Z2t where we carefully remove the most insignificant
terms, and we end up with:

D4Invt = µ1 + Π1,1 D4Invt−1 + Π4,1 D4Invt−4 + Π5,1 D4Invt−5 + Π8,1 D4Invt−8 + Π9,1 D4Invt−9 + Π1,2 D4Qt−1
+ Π2,6 D4Qt−6 + θ1,1 D2000(1) + ϵ1t
(11)
D4Qt = µ1 + Π6,1 D4Invt−6 + Π1,2 D4Qt−1 + Π4,2 D4Qt−4 + Π5,2 D4Qt−5 + Π6,2 D4Qt−6 + Π8,2 D4Qt−8
+ Π9,2 D4Qt−9 + θ1,2 D2000(1) + ϵ2t
(12)

6
 Table 1: Estimation results VAR.

Variable\Method D4Inv Robust SE D4Q Robust SE

D4Inv1 0.920 (0.045) - -
D4Inv4 -0.671 (0.087) - -
D4Inv5 0.649 (0.0.89) - -
D4Inv6 - - 0.795 (0.336)*
D4Inv8 -0.406 (0.075) - -
D4Inv9 0.276 (0.058) - -
D4Q1 0.013 (0.004) 0.995 (0.044)
D4Q4 - -0.576 (0.082)
D4Q5 - - 0.612 (0.093)
D4Q6 -0.01 (0.003) -0.204 (0.099)
D4Q8 - - -0.216 (0.112)
D4Q9 - - 0.153 (0.082)
dummy20001 0.009 (0.000) 0.012 (0.002)
Constant -0.0001** (0.000) 0.003** (0.001)
D4Qt -0.0031** (0.01454) - -
Correlation D4Inv 0.002 0.357
Correlation D4Q 0.357 0.027
σ 0.021 0.027
N 182 182
Notes: The values in parentheses are the robust standard errors (HCSE).

As we have 9 lags, our estimation starts in 1974(3) and ends in 2019(4).

Equation Excluded df Statistic

Q4Inv D4I2 , D4I3 , D4I6 , D4I7 , D4Q2 , D4Q3 , D4Q4 , D4Q5 , D4Q7 , D4Q8 , D4Q8
D4Q D4I1 , D4I2 , D4I3 , D4I4 , D4I5 , D4I7 , D4I8 , D4I9 , D4Q2 , D4Q3 , D4Q7 − −

χ2 − 22 17.223

From the Wald-test it can be derived that the model of our choice would be model Z2t , as we can
reject that there would be an omitted variable bias. The estimation of Z2t does not show any auto-
correlation, we then estimate the root of the companion matrix graphically to show that the model
is stationary.

7
 Figure 4: Root of the companion matrix plot

The eigenvalues are inside the complex unit circle and this indicates that we have a consistent
and stable estimator.

A change in the first and sixth periods of Tobin’s Q has a positive effect on the residential investment,
where the effects might not be sizeable they are significant. A change in the sixth period of residential
investment has a sizable positive, and significant3 effect. As there is an apparent contemporaneous
effect, our approach to deducing the direction of the causality will consist of the no-Granger causality
test.

As we have heteroskedasticity, we do a robust Wald-test to test for the hypothesis of no-Granger

causality in both directions, D4Qt ↛ D4Invt , D4Invt ↛ D4Qt .

Equation Excluded χ2 df Prob >χ2

Q4Inv D4Q1 D4Q6 23.088∗∗ 2 0.00
D4Q D4Inv6 3.566 1 0.05
From the test statistics, it can be seen that we reject the null hypothesis of no Granger causality in
the direction from Tobin’s Q to residential investment. Altogether we may conclude that the Granger
causality runs from Tobin’s Q to residential investment, and not the other way around. The result
of the test coincides with the macroeconomic theory.

For further analysis of the dynamic relationship, we assume that the causality between the vari-
ables runs from Tobin’s Q to residential investment. This opens the possibility for us to analyze how
long the effect of a change in Tobin’s Q impacts the residential investments, by defining the structural
VAR model :

D4Invt = µ1 + Π1,1 D4Invt−1 + Π4,1 D4Invt−4 + Π5,1 D4Invt−5 + Π8,1 D4Invt−8 + Π9,1 D4Invt−9 + Π1,2 D4Qt−1
+ Π2,6 D4Qt−6 + θ1,1 D2000(1) + ΦD4Qt + ϵ1t
(13)
D4Qt = µ1 + Π6,1 D4Invt−6 + Π1,2 D4Qt−1 + Π4,2 D4Qt−4 + Π5,2 D4Qt−5 + Π6,2 D4Qt−6 + Π8,2 D4Qt−8
+ Π9,2 D4Qt−9 + θ1,2 D2000(1) + ϵ2t
(14)
3 On a 10% significance

8
From Table 1, it is clear that the contemporaneous effect is not instantaneous, as the effect of Tobin’s
Q is not significant in the same quarter. To capture the dynamic relationship, a feasible option for
us is to analyze the impulse-response:

Figure 5: Impulse responses (orthogonalized)

The above graphs reflect the impulse response functions for our VAR model. By looking at the
top left graph we can see that a shock variable D4 Inv has a positive impact itself. We also see the
same relation in the bottom right example with D4 Q. More interestingly when looking at the bottom
left graph (shock to D4 Q impact on D4 Inv), we see that the cumulative impact converges to around
0.0005 over time. This suggests that Tobin’s Q does not impact the result of the real residential
investment. Finally looking at the reverse relationship, we see that shock to D4 Inv has a positive
cumulative effect on D4 Q indicating that Tobin’s Q is somewhat dependent on the real residential
investment.

5 Discussion
From our empirical analysis, it has been derived that the dynamic relationship coincides with the
macro-economical theory that movements in Tobin’s Q drive residential investments. From our analy-
sis, it can be seen that the change of Tobin’s Q does not have an instantaneous effect on the residential
investment, but the effect is captured around the 6th quarter. It is also important to note, that even
though there might be a causality, the effects are limited and minor. From our analysis, it is worth
pointing out, that the sixth residential investment also has an impact on Tobin’s Q. It is more evident,
that there are omitted variables, such as mortgage rates and consumer confidence which might have
a higher impact on residential investment and more prompt.

6 Conclusion
After analyzing the dynamic relationship between changes in residential investments and Tobin’s Q,
we find that movements in Tobin’s Q drive residential investments as the theory suggests. The effects
are small but they are significant, this indicates that other factors may have a larger influence.