Appendix II

1. INTRODUCTION
The statistical test and concepts used in various chapters of this thesis could not be discussed in detail
due to space considerations. We hereby note done down the basis of some of the important statistical
tests that have been used time and again in this thesis.

2. UNIT ROOT TEST

2.1 The Augmented Dickey-Fuller (ADF) Test

To illustrate the use of Dickey-Fuller tests, consider first an AR(1) process:

where p and p are parameters and st is assumed to be white noise, y is a stationary series if -l<p<l.
If p=l, y is a nonstationary series (a random walk v/ith drift); if the process is started at some point,
the variance of y increases steadily with time and goes to infinity. If the absolute value of p is
greater than one, the series is explosive. Therefore, the hypothesis of a stationary series can be
evaluated by testing whether the absolute value of p is strictly less than one. Both the DF and the PP
tests take the unit root as the null hypothesis Ho: p=l. Since explosive series do not make much
economic sense, this null hypothesis is tested against the one-sided alternative Hp pel.

The test is carried out by estimating an equation with yt.j subtracted from both sides of the equation:

4f,=M+'9,-i+e<
where v = p-1 and the null and alternative hypotheses are
H0:v = 0, Hi:v<0

While it may appear that the test can be carried out by performing a t-test on the estimated v, the t-
statistic under the null hypothesis of a unit root does not have the conventional t-distribution.
Dickey and Fuller (1979) showed that the distribution under the null hypothesis is nonstandard, and
simulated the critical values for selected sample sizes. More recently, MacKinnon (1991) has
implemented a much larger set of simulations than those tabulated by Dickey and Fuller. In
addition, MacKinnon estimates the response surface using the simulation results, permitting the
calculation of Dickey-Fuller critical values for any sample size and for any number of right-hand
variables. EViews reports these MacKinnon critical values for unit root tests.

The simple unit root test described above is valid only if the series is an AR(1) process. If the series
is correlated at higher order lags, the assumption of white noise disturbances is violated. The ADF
and PP tests use different methods to control for higher-order serial correlation in the series. The
ADF test makes a parametric correction for higher-order correlation by assuming that the y series
follows an AR(p) process and adjusting the test methodology. The PP approach is described below.
85
 The ADF approach controls for higher-order correlation by adding lagged difference terms of the
dependent variable y to the right-hand side of the regression:

Ay, =//+vy,_t +<M.y,-x +<M.y(-2 +...................+SpAy,^, + e,

This augmented specification is then used to test:

H0: v = 0, H,:v<0
in this regression. An important result obtained by Fuller is that the asymptotic distribution of the t-
statistic on v is independent of the number of lagged first differences included in the ADF
regression. Moreover, while the parametric assumption that y follows an autoregressive (AR)
process may seem restrictive, Said and Dickey (1984) demonstrate that the ADF test remains valid
even when the series has a moving average (MA) component, provided that enough lagged
difference terms are augmented to the regression.

2.2 The Phillips-Perron (PP) Test

Phillips and Perron (1988) propose a nonparametric method of controlling for higher-order serial
correlation in a series. The test regression for the Phillips-Perron (PP) test is the AR(1) process:
Ay, = a+{fy,_t +e,

While the ADF test corrects for higher order serial correlation by adding lagged differenced terms
on the right-hand side, the PP test makes a correction to the t-statistic of the y coefficient from the
AR(1) regression to account for the serial correlation in s. The correction is nonparametric since
an estimate of the spectrum of e at frequency zero that is robust to heteroskedasticity and
autocorrelation of unknown form is used. Newey-West heteroskedasticity autocorrelation
consistent estimate

where q is the truncation lag. The PP t-statistic is computed as

t rlnt> W-nK
pp <o 2 <»<x
where h and Ob are the t-statistic and standard error of p and a is the standard error of the test
regression.

The asymptotic distribution of the PP t-statistic is the same as the ADF t-statistic which reports
MacKinnon critical values. As with the ADF test, one can specify whether to include a constant, a
constant and linear trend, or neither in the test regression. For the PP test, one has to specify the
truncation lag q for the Newey-West correction, that is, the number of periods of serial correlation
to include.

86
3. JOHANSEN’S COINTEGRATION TEST

3.1 Test Specification

Given a group of non-stationary series, the Johansen’s Test determines whether the series are
cointegrated, and if they are, in identifying the cointegrating (long-run equilibrium) relationships.
Johansen (1991, 1995) developed a methodology of Vector Autoregression (VAR) - based
cointegration tests which is to test the restrictions imposed by cointegration on the unrestricted VAR
involving the series.
Johansen’s Cointegration Test considers a VAR of order p in the form:

y. =Aiy^i+......... +APy,-P+Bxt+sf

where yt is a k-vector of non-stationary 1(1) variables, xt is a d-vector of deterministic variables, and st

is a vector of innovations. The VAR can be rewritten as:
p-i
Ay, =nyt_1 + 2]ri Ayt_, + Bxt + et
i=l

where n = ^Aj-Ii and T; =-^Aj

i=! j=i+l

Granger’s representation theorem asserts that if the coefficient matrix n has reduced rank r<k, then there

exist k x r matrices a and p each with rank r such that ^l = aP and ^y, js stationary, r is the number of

cointegrating relations (the cointegrating rank) and each column of P is the cointegrating vector. The
elements of a are known as the adjustment parameters in the vector error correction model. Johansen’s
method is to estimate the k matrix in an unrestricted form, then test whether we can reject the
restrictions implied by the reduced rank of n.

3.2 The number of cointegrating relations

If there arek endogenous variables, each of which has one unit root, there can be from zero to k-1
linearly independent, cointegrating relations. If there are no cointegrating relations, standard time series
analyses such as the (unrestricted) VAR may be applied to the first-differences of the data. Since there
are k separate integrated elements driving the series, levels of the series do not appear in the VAR in
this case.

Conversely, if there is one cointegrating equation in the system, then a single linear combination of the
levels of the endogenous series P'yt-i, should be added to each equation in the VAR. When multiplied by
a coefficient for an equation, the resulting, term a p'yt-n is referred to as an error correction term. If there
are additional cointegrating equations, each will contribute an additional error correction term involving
a different linear combination of the levels of the series.

If there are exactly k cointegrating relations, none of the series has a unit root, and the VAR may be
specified in terms of the levels of all of the series. Note that in some cases, the individual unit root tests

87
will show that some of the series are integrated, but the Johansen tests show that the cointegrating rank
is k. This contradiction may be the result of specification error.

3.3 The Cointegrating Relations (Vector)

Each column of the J3 matrix gives an estimate of a cointegrating vector. The cointegrating vector is not
identified unless we impose some arbitrary normalization. EViews adopts the normalization so that the r
cointegrating relations are solved for the first r variables in the yt vector as a function of the remaining
k-r variables.

Note that one consequence of this normalization is that the normalized vectors which EViews provides
will not, in general, be orthogonal, despite the orthogonality of the unnormalized coefficients.

3.4 Deterministic Trend Assumptions

Your series may have nonzero means and deterministic trends as well as stochastic trends. Similarly, the
cointegrating equations may have intercepts and deterministic trends. The asymptotic distribution of the
LR test statistic for the reduced rank test does not have the usual %2 distribution and depends on the
assumptions made with respect to deterministic trends. EViev/s provides tests for the following five
possibilities considered by Johansen (see Johansen, 1995, pp. 80-84 for details):

1. Series y has no deterministic trends and the cointegrating equations do not have intercepts:
H2(r): nj-,-, =
2. Series y has no deterministic trends and the cointegrating equations have intercepts:

(r): II+ B/2'y,_t = a( fi'y,^ + p<>)

3. Series y has linear trends but the cointegrating equations have only intercepts:
Hx(r): ]J v.-i + Bfiy,_, = a( P'y,_x +p0) + a±y„
4. Both series y and the cointegrating equations have linear trends:
Hl(r): Uy^+Bp'y,^ = a(P'y,_1+p()+pit)+a1ro

5. Series y has quadratic trends and the cointegrating equations have linear trends:
H(r): = a(P'y,-i +P„ +Pit)+a±(Yo + 7iO

Where, a± is the (non-unique) k x (k-r) matrix such that a'a± =0 and rank flor| ax |)=k .

These five cases are nested from the most restrictive to the least restrictive, given any particular
cointegrating rank r:
H1(r)czH'1(r)c:Hl(r)c:H'(r)cH'(r)

For each case, EViews tabulates the critical values for the reduced rank test as given by Osterwald-
Lenum (1992), not those tabulated in Johansen and Juselius (1990). Note that the critical values are
available for up to 10 series and may not be appropriate for models that contain other deterministic
regressors. For example, a shift dummy variable in the VAR implies a broken linear trend in the y
series.
88
 4. VECTOR ERROR CORRECTION MODEL

Consider two random walk processes {Yt} and {Xt}, that is, they are integrated of order one, 1(1).

Yt=YM+et
X, =xt_, +£t
where £, is a stationary process or 1(0). Let us specify a regression model without the intercept term as
shown below:
Yt = pX, +ut ..... (1)

Now if there is a non-zero (3 such that Yt _PXt is a 1(0) process, i.e. is a 1(0), then say that Yt and

Xt are cointegrated. Then an equation in the first difference of the form:

AY, = ocAXt +A,(Yt-pXt) + v, .... (2)

is a valid equation. Since AYt, AXt, (Yt - pxt) and vt are all 1(0). Equation (1) is considered a long run
relationship between Yt and Xt and equation (2) describes the short run dynamics. Engle-Granger

suggest estimating (1) by ordinary least squares to obtain the estimate P P, which then is to be
substituted in (2) to estimate the parameters a and X. This two step estimating procedure rests on the
assumption that Yt and Xt are cointegrated and thus it is important to test for cointegration. Engle-

Granger suggest estimating (1) by ordinary least squares, getting the residuals u‘ and then applying the

Dickey-Fuller test based on u'. Engle-Granger suggests that to test cointegration for the residuals, the

ADF critical values are to be adopted. Thus, Engle and Granger (1987), pointed out that, a linear
combination of two or more non-stationary series may be stationary. If such a stationary, or 1(0), linear
combination exists, the non-stationary (with a unit root), time series are said to be cointegrated. The
stationary linear combination is called the cointegrating equation and may be interpreted as a long-run
equilibrium relationship between the variables. The VEC specification restricts the long-run behavior of
the endogenous variables to converge to their cointegrating relationships while allowing a wide range of
short-run dynamics. The cointegration term is known as the error correction term since the deviation
from long-run equilibrium is corrected gradually through a series of partial short-run adjustments.

89