Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Aggregate Import Demand function for Japan: A

cointegration approach
Razvan Cucu
University of Brighton, Brighton, UK

Introduction

The purpose of this econometric paper is to examine the long run stability of the import
demand function for Japan, and whether the short-run model can indeed adjust/correct the
long-run if necessary. In addition, it will identify to what extent the determinants of import
demand affect its short- and long-run equilibrium. The method used will be a Cointegration
technique advocated by Engle and Granger in 1987, including the error correction model, in
order to create a dynamic short-run model.
Therefore on cointegrating relation investigation is necessary along with the long-run
elasticities estimates, in order to draw a solid conclusion in regards to Japan‟s aggregate
import demand behaviour, and whether there is a stable long-run equilibrium. This requires
an in-depth investigation on the existence of conintegration between the real imports of Japan
and the deterministic variables: real national income represented by the GDP and the cost of
imports represented by the ¥/$ exchange rate. This will give us a clear idea of the long-run
equilibrium relation in which the dependent variables will not move too far apart, if so, will
be adjusted back to their equilibrium by the short-run error correction model (Tang, 2008).
Therefore this will enable us to obtain the not only the level of cointegration between the
determinant variables, but also the stability of the import demand function; accounting for the
main purposes of this academic paper.
Previous empirical studies have been preoccupied with the cointegration method applied
on import demand, such as Meese and Singleton (1983) and Nelson and Plosser (1982) have
stated that „aggregate economic time series are not stationary in their levels and therefore
contain variances that explode with time‟. This proves (Fuller, 1985) statements which says
that because limited distribution of the asymptotic variance in the time-series might result in
misleading results, especially on small sampling intervals of data, which is applicable in our
case. Therefore if the investigation does not account for unit root non-stationarity, then our
model can suffer from i) spurious regression issue (Phillips, 1986); ii) inconsistency and
inefficiency in the OLS estimations, unless variables are cointegrated and iii) we cannot
create a ECM in order for adjustment(Engle and Granger, 1987). We will investigate the
issues in terms, and add our own unique participation, with the purpose of stating a solid
evidence based conclusion.
The aggregate import demand model

The standard form for the import demand function in macro-economy is based on the
fundamental assumptions that import demand is determined mainly by the national income
and the price of imports. There are of course other determinant factors that could be included,
but as Hong (1999) mentioned, they all can be subsumed within these two primordial factors,
at least from a theoretically perspective. Assuming we have zero degree homogeneity and
infinite supply elasticity, we can formulate the import demand function with the following
equation:
𝑀𝑡 = 𝑓(𝑌𝑡 , 𝑃𝑡 )
This suggests that the desired quantity of import demanded 𝑀𝑡 is a function of the real
national income 𝑌𝑡 in domestic currency, and the price of imports 𝑃𝑡 . From this equation we
can now derive our log-linear regression model for the import demand function:

𝑙𝑛𝑀𝑐𝑡 = 𝛽0 + 𝛽1 𝑙𝑛𝑌𝑐 + 𝛽2 𝑙𝑛𝐸𝑟 + 𝑢𝑡 where 𝑢𝑡 ~NID(0, 𝜎 2 )

The error term assumed to have the following properties:

A1. E(ui) = 0 [zero mean]


A2. V(u) = 2 [constant variance]
A3. E(uu‟) = 0 [zero autocorrelations]
A4. E(Xu) = 0 [X is is non stochastic]

Written in natural logarithmic form, Mc represents the quantity of imports at period t, Yc


represents the real GDP at period t and Er represents the ¥/$ exchange rate. 𝛽0 represents the
linear constant and 𝑢𝑡 a random error term (residuals). Economic theory suggests that Yc
should be positively correlated to Mc, therefore an increase in income should result with an
automatic increase in the quantity of imports demanded; whereas Er should have a negative
relationship to Mc, suggesting an increase in Er should force the imports demand to fall,
based on the Keynesian perspective. This expects that Yc to have positive signs, and Er
negative signs in our estimated model.

Data
The dataset used for this model is a collection of time-series data from OECD, for the period
1950-1990 therefore we have 41 observations. We have selected a number of economic
indicators, used as variables in concordance with the economic theory. Firstly, we have our
dependant variable „Import demand‟, noted as „mc‟ which represents imports at current
prices. Secondly we have our two independent variables; Income, noted as „yc‟ representing
the GDP at current prices, and Price of imports, noted as „er‟ which represents the ¥/$
exchange rate. Although economic theory suggests that there might be other determinants for
the demand of imports, we have included them in addition to the first three. These variables
tested were: Government expenditure, Investment, Employment and Unemployment but they
all proved statistically insignificant (at all significance levels, standard errors and prob.
values), and confirmed using the data, that they do not affect the import demand function.

The reason we have chosen „mc‟, „yc‟ and „er‟ as our variables are based on the theoretical
model of the aggregate import demand which states that the quantity of imports (Md) is
negatively related to the to the relative price of imports (P), and positively to the real national
income (Y); which could be illustrated in a linear equation as follows:𝑀𝑑 = 𝑃 + 𝑌

There have been a series of transformations to our dataset, in order to for the data to be
statistically relevant and significant, converting it into the same units, account for
measurements errors and use proxy variables if necessary; but also for qualitative reasons
such as being unproblematic to interpret or to disentangle the effects of an independent
variable on the dependant one. First transformation applied is the deflation of all our
variables in order to get our values in real terms rather than nominal by removing the effects
of inflation over time, which could create biased assumptions and propositions especially for
economic time-series data. The method used was simply divided each variable by the retail
price index and saving it as a new variable.

Subsequent to that stage, another issue arisen, which was that our data was measured in
different units and often a log-linear transformation is necessary especially with time-series
data in order to induce symmetry. Therefore a log-linear form of equation was preferred to a
linear-form in order to provide a constant elasticity regression in levels. The transformation
took place with a simple replacement of the original variables, with the new variables created
in logarithmic form, created in eviews: 𝑙𝑛𝑀𝑐 = log 𝑀𝑐 ; 𝑙𝑛𝐸𝑟 = log 𝑒𝑟 ; 𝑙𝑛𝑌𝑐 =
log⁡(𝑦𝑐). These three new variables simply represent elasticities of the original variables,
rather than actual values, making it easier to model and interpret. After these transformations
our economic model estimation for the import demand function is:

𝐼𝑚𝑝𝑜𝑟𝑡 𝑑𝑒𝑚𝑎𝑛𝑑𝑡 = 𝛽0 +𝛽1 𝐺𝐷𝑃𝑡 + 𝛽2 𝐸𝑥. 𝑟𝑎𝑡𝑒𝑡 ,


and our log-linear regression estimation for import demand is:

𝑙𝑛𝑀𝑐𝑡 = 𝛽0 + 𝛽1 𝑙𝑛𝑌𝑐 + 𝛽2 𝑙𝑛𝐸𝑟 + 𝑢𝑡

Methodology
The first test conducted on our model is the t-test on each specific coefficient. The results
show correct signs (positive) on the Yc variable in relation to Mc, meaning that 1% increase
in income should boost import demand by 1.10%; and the prob. value is <0.05 which
suggests that we accept the null hypothesis, meaning that this variable has no direct effect on
the dependent variable. For the Er variable the signs are incorrect (positive instead of
negative) which disagrees with economic theory which suggests that exchange rates should
be negatively related to import demand. Although the prob. value is <0.05 which means it is
significant. We now have to correct it, and the method which proved most effective is by
adding two dummy variables in the equation. The dummy variables, both for the Er,
accounting for i) structural change in the Japanese economy and trade balance adjustments,
and ii) for the switch from a fixed exchange rate system to a floating exchange rate one. After
running the regression with the two dummy variables included, the Er showed correct signs
and significance, therefore correcting our model. The dummy variables also had correct signs
and prob. values were <0.05 which proved their significance level.
The second test we have looked at was the F-test statistic, which tests for joint
significance of all included variables; in our case we had a high F-stat (1089.9) and prob.
value<0.05 which suggests we accept the null hypothesis that all coefficients (except
constant) are zero simultaneous. In order to test for the goodness of fit of our estimated
model, we analyse the 𝑅 2 value, in our case it is (0.993) and adjusted 𝑅 2 (0.992), which
clearly show that we have a good fitted model (close to 1 rule).
In order to test for correct functional form we have ran the stability Ramsey RESET Test
on the fitted square values, and the F-stat prob. value was (0.96) which clearly indicates that
we accept the null hypothesis of correct functional form, and therefore our model is correctly
specified.
We have then tested for normal errors using the Jarque-Bera normality test on our
residuals. The Jarque-Bera value was (0.14) and prob. value (0.93); we accept the null
hypothesis that the errors are normally distributed based on the rejection rule. Although the
histogram for the residual plot doesn‟t fully resemble normal distribution we still accept the
null, supported by theory which suggests that the Jarque-Bera is an asymptotic test and it‟s
mostly applicable to large samples of data in order for it to be totally valid. (Mukhurjee et al.
1998)
Based on an auxiliary regression we have tested our model for heteroscedasticity using
the White‟s test (residual test). The χ2 value was (0.026) which suggest that we reject the null
of homoscedastic error variance, based on the <0.05 critical value. In order to correct for
heteroscedasticity I have included a @trend variable as suggested in the literature, which
proved efficient. After running the White‟s test with the @trend included in the equation, the
χ2 value was (0.54) which clearly suggests we accept the null hypothesis of homoscedastic
error variance.
In order to test for autocorrelation we use, the Breusch-Godfrey test using an auxiliary
regression, which tests for serial independence in the error terms, and also tests for higher
orders of autocorrelation than AR(1), unlike the Durbin Watson test. The χ2 prob. value is
(0.134) which suggests we accept the null hypothesis of serial independence, meaning there
is no autocorrelation in our residuals. Just as a 2nd method, we have also implemented the
Durbin-Watson test for precautionary measures, which also proves no autocorrelation by the
rejection rule: DW>DW𝑈 || 1.92>1.42 for 45(41 not on table) observations and three
regressors excluding the constant.
In order for our data to be stationary it must comply with the following assumptions:

T1. 𝐸(𝑦𝑡 ) = 𝜇 [constant mean]

T2. 𝐸[(𝑦𝑡 − 𝜇)2 = 𝑣𝑎𝑟(𝑦𝑡 ) [constant variance]

T3. 𝐸 (𝑦𝑡 − 𝜇 (𝑦𝑡−𝜏 − 𝜇)] = 𝑐𝑜𝑣(𝑦𝑡 , 𝑦𝑡−𝜏 ) [autocovariances independent of time]

By looking at our data in a line graph we visually identify that there is a pure random walk,
statistically denoted as:
𝑦𝑡 = 𝜌𝑦𝑡 + 𝑒𝑡 where 𝑒𝑡 ~𝑖𝑖𝑑(𝑜, 𝜎 2 )
We then visually inspect the correlograms for all the variables and they illustrate a slow
decay (die out slowly) of autocorrelation between the neighbouring data points which
suggests that our time-series data is non-stationary, and therefore concludes a pure random
walk. This means that our data satisfies T1, but not T2 nor T3. But since the correlogram is
only visually inspect, therefore it is not an exact instrument to identify stationarity, we will
use the Dickey-Fuller test in order to test for unit root.
First of all we have tested all our variables for unit root at level with intercept and drift
and the Dickey-Fuller t-stat for each variable were: lmc(-1.63), lyc(-0.22), ler(-1.65). These t-
stats do not fall under any of the critical values beneath them, which means there is a unit
root present in all our variables, therefore we accept the null hypothesis of non stationarity.
Due to the fact that we have failed to reject the null hypothesis in levels, we now test our data
for unit root in 1st difference; the result showing the following t-stats: lmc(-5.18), lyc(-2.92),
ler(-4.36) all with the prob. value <0.05. In this case we can reject the null hypothesis and
accept the alternative, which states that the time-series are stationary and integrated at order
one I(1).
Because the Dickey-Fuller test only tests for the 1st order of autocorrelation I(1), so if
hypothetically the order is higher, the test becomes invalid and the DF equation its affected
from residual correlation and unit root. Therefore we run an Augmented Dickey-Fuller test,
which augments the Dickey-Fuller equation by lagged values to the dependent variable, in
order to test that the estimation equation is not residually correlated (white noise) and also
that the dependent variable lmc is not characterised by an higher order autoregressive
process; using lag length automatically selected by eviews based on the Schwarz Info
Criterion as follows for each variable. Subsequently, having ran the ADF test and the t-stats
are as follows: lmc(-4.25); lyc -4.30; ler-3.76) all with prob. values <0.05 which suggests that
straight forward we can reject the null hypothesis and accept the alternative of stationary
time-series integrated of order 1 I(1).
Empirical Results
In order to correct our data for non-stationarity, as its integrated at I(1), we have to
difference every single variable apart of the two dummy variables and @trend. The method
of doing this process is by creating a set of new variables as follows:
𝛥𝑙𝑚𝑐 = 𝑙𝑚𝑐 −1 ; 𝛥𝑙𝑦𝑐 = 𝑙𝑦𝑐 −1 ; 𝛥𝑙𝑒𝑟 = 𝑙𝑒𝑟(−1) .

We now save the new created variables as dmc; dyc; der, which can give a measure of
long-run elasticities.

Using the Engle-Granger two-stage procedure, we now run the unit root test (ADF) on them
in levels to check if non-stationarity I(1) was corrected. The new t-stats for each variable
show: dmc(-4.10); dyc(-4.30); der(-4.96), all with prob. values <0.05. This has proved that by
taking the difference on each variable has corrected our model for non-stationarity, and now
we can now reject the null hypothesis and state that our series is integrated at order zero I(0)
and stationary. Subsequently, we retrieve the newly estimated residuals and check them for
stationarity. The ADF t-stat for the residuals is (-4.78) which suggests that it is stationary.
Since they are stationary, it can be interepreted as there is a long-run relationship between our
variables. Due to near singular matrix we have decided to remove the dummy variables from
the short run equation, as it wouldn‟t allow it to run with the dummies included, probably due
to collinearity. We then save the residuals as ECM (ecm= resid), and run the Error Correction
Model dynamic equation of a short-run disequilibrium:
𝛥𝑚𝑐𝑡 = 𝛽0 + 𝛽1 𝛥𝑦𝑐𝑡 + 𝛽2 𝛥𝑒𝑟𝑡 + 𝑒𝑐𝑚𝑡−1

The ecm coefficient is (-5.62) and <0.05 significant; having the right sign (negative) and
prob. value <0.05 significance. This coincides with economic theory as it reflects the speed of
adjustment to the long-run equilibrium, although the duration of adjustment may be slightly
longer. This proves that our model is convergent.

We now present our long-run static model from which we obtain our long-run elasticities (*
represents significance levels, st. errors in parenthesis):

lmc= 2.255 +1.099***lyc –0.903***ler –0.764***dummy –0.445***dummy2– 0.019*@trend


(1.552) (0.190) (0.187) (0.110) (0.076) (0.025)

And our short-run dynamic model with the ecm incorporated as mentioned above:

Δlmc = 1.470 + 0.920**Δlyc – 0.579*Δler – 5.622*ecmt−1


(6.356) (0.290) (0.763) (2.873)
Empirical Analysis and interpretation

We can clearly state that all the signs are correct on both the long-run and the short-run
model, although significance levels are stronger in the long-run which shows a stronger
affects the dependent variables have on the import demand in the long-run equilibrium, as
well as it may take a while for the changes in national income and exchange rate fluctuations
in order to have a significant impact on the import demand, in the long run. In the short-run
we can identify that changes in national income has a stronger impact on the imports demand,
than the level of the exchange rate. This could be supported by macro-economic theory by
stating that an increase in GDP would encourage the population to demand more imports as
their expendable income also rises, whereas small fluctuations in the exchange rate would
have a lesser effect.
For a direct interpretation of our empirical results we can state that in the long-run
equilibrium, a 1% increase in national income (GDP) raises the demand for imports by 0.90%
which shows that it is a significant determinant as theory suggests. Whereas a 1%
depreciation of the Yen in regards to USDollar, decreases the import demand by 0.76%,
although highly significant. Therefore in the long-run both dependant variables have a fairly
similar determinacy level. This somewhat changes in the short-run dynamic equilibrium,
where a 1% increase in the GDP raises the demand for imports by 0.92%, whereas 1%
depreciation of the ¥/$ only decreases the import demand by 0.58%, showing a weaker affect
than in the long-run. This could be interoperated as the short run level of fluctuations in the
exchange rate is relatively insignificant(white noise) compared to the increase in national
income which is fairly larger proportions.
The ECM is stationary in regressors, although the coefficient for ecm (-5.622) indicates
it may take a longer period for the short-run to adjust the long-run equilibrium when
necessary, but we can strongly state that it will correct it in the long-run, based on the
negative sign of the coefficient.

Conclusion

Using the aggregate import demand function, with the aid of cointegration and error
correction model, we have found that a long-run equilibrium relationship exists between the
deterministic variables. Although for the short-run ECM would take a while in order to fully
adjust the disequilibrium, based on the empirical evidence which suggests that both our
model are robust and statistical significant with minor discrepancies under the diagnostic
tests. With the use of dummy variables along the elementary ones (GDP and Exchange rate)
we have determined the extent of impact that each of these have on the level of import
demand over two time frames (short- and long-run). This could be related to the work of
Hamori and Matsubayashi (2001), in which they have said that in order to ensure a stronger
long-run equilibrium, Japan should reduce its trading surplus in order to stimulate the
domestic demand by increasing imports and therefore reducing the trade surplus.
Subsequently, further improvements of our model could be made in the future, in regards to
additional variables, larger dataset to account for structural change, time-volatility concept
and substitutions of different variables for other types of measurement.

References:

Engle, R. F. and Granger, C.W. J. (1987) Cointegration and error correction: representation,
estimation and testing, Econometrica

Fuller, W.A. (1985) Nonstationary autoregressive time series, Handbook of Statistics, 5, North-
Holland, Amsterdam

Hamori, S. and Matsubayashi, Y., 2001, An empirical analysis on the stability of Japan‟s aggregate
import demand function, Japan and the World Economy, 13

Hong, Pingfan, 1999, Import Elasticities Revisited. Department of Economic and Social Affairs,
Discussion Paper No. 10, United Nations

Meese, R. A. and Singleton, K. J. (1983) On unit root and the empirical modelling of exchange rates,
International Economic Review

Nelson, C. R. and Plosser, C. I. (1982) Trends and random walks in macroeconomic time series,
Journal of Monetary Economics

Phillips, P. C. B. (1987) Understanding spurious regression in econometrics, Journal of Econometrics

Tuck Cheong Tang, 2008, “AGGREGATE IMPORT DEMAND FUNCTION FOR JAPAN: A
COINTEGRATION RE-INVESTIGATION”, Asian Business and Economics Research Unit
Discussion Paper 63, 2008

=38 + 298

You might also like