TH E ESTIMATION PROCEDURE : UNIVARIATE MODELLING

(4.9c)

The main focus of this study is to investigate the impact of FDI on economic growth, exports, imports, capital formation and productivity and their causality during the period of study by employing annual time-series data for the period 1 960 to 200 1 .

In general, classical economic theory assumes that observed data come from a

stationary process, where means and variance are constant over time. 10 Granger and

Newbold ( 1 974) demonstrated that most time series variables share common trends and

have the potential to reveal spurious results. Therefore, when applying the estimation

techniques to time series data, some model builders make stationary assumptions about

9 T here i s no unique measure of openness to trade, see Edwards, ( 1 998) for a succinct discussion on various measures of openness.

1 0

Refer to Chapter Six for details.

the variables used in the models. However, graphs of economic time series reveal the invalidity of such assumptions, and led to the analysis of unit-root processes and cointegration. In recent times, unit-root processes and cointegration systems have played a prominent role in econometrics and macroeconomics, with applications to such diverse fields as fmance, economic history, international trade, and so on. The reasons behind such an expansion are the strong intuitive appeal and highly involved technical complexity.

A non-stationary process arises when one of the above mentioned assumptions for stationarity does not hold. Therefore, cointegration analysis has been extensively applied to test the long-run equilibrium relationships among non-stationary economic variables. The time series of several variables, X;, are co integrated if these variables are

individually nonstationary but there exists at least one linear combination of them, Zt =

/3 'X;, that is stationary. Such cointegrated variables do not drift far apart and they tend to

move together in the long run. Thus, when applying time series data in empirical analysis researchers should pay attention to the stationarity and cointegration properties of data.

To tackle the aforementioned problems, this study, therefore uses the ARDL method recently developed by Pesaran and Shin ( 1 995) 1 1 to investigate the impact of FDI on the individual variables proposed. In addition, the study proposes to utilise the cointegration technique, which has received much attention since Granger formally introduced it in 1 98 1 . A simple illustration of the ARDL technique using a model with two variables and a maximum of two lags is explained in Appendix 4.2.

Furthermore, the main concern of this study is to investigate FDI's impact on key macroeconomic variables. As FDI does not appear only as a one-time effect on the host country's investment level, but results in long-term effects, some researchers, when modelling FDI influence on other key macroeconomic variables, apply lagged forms of 1 1 _ARDL

applications and econometrics theory are analysed in Pesaran and Shin ( 1 995) and Pesaran and Pesaran ( 1 997).

the dependent and independent variables as explanatory variables. These can be seen in the ARDL procedure because ARDL models are dynamic in nature and explicitly consider the behaviour of a variable over time. The model shows how a change in an explanatory variable affects the dependent variable, which is "distributed" over a measurable number of future time periods. This enables the ARDL modelling procedure to test for short- and long-run relationships between sets of variables (Hendry, 1 995). Short-run relationships between variables do not persist over long periods, making it possible to detect temporary disturbances to the links between variables, which can be picked up in the regressions. Likewise, the long-run relationship is useful in assessing shocks that occur over time, can be used regardless of the order of integration of the variables, and avoids pre-testing problems associated with other co integration methodologies. Consequently, the ARDL technique minimises the possibility of spurious relation through non-stationary data while retaining valuable long-run information (Hendry, 1 995). The system has an augmentation procedure that uses the minimum number of lags on each variable consistent with statistical significance to remove serial correlation from the error terms (Pesaran and Pesaran, 1 997), advantages that justify the estimation procedure adopted in this study.

The ARDL procedure consists of two stages. In the fust stage, the existence of a long­ run relationship between variables concerned is examined by computing the F-statistic to test the significance of the lagged levels of variables of the series in the error

correction form of the underlying ARDL model (Karfakis and Phipps, 200 1 ). The test

statistic is a joint test of the null hypothesis that coefficients Co and C I in the following

equation (4. 1 0) equal zero, i.e. testing whether lagged levels of variables X and Z are

jointly i nsignificant. T hus H o: C l = C o = o . Rejection 0 f H 0 i mplies t he e xistence 0 f

long-run relationships between the variables. The error correction version of the ARDL

(p,q) in the variables Xt and Zt is given by:

p q

Mr

I A;L:1Xt-(

I B/:1Zr_j

Jir

i=t

j=t

The computed F-statistics under Ho: Cl = Co = 0 have a non-standard asymptotic

distribution regardless of the integration of the variables. Pesaran and Pesaran ( 1 997)

. provide the tabulated appropriate critical values,12 which consist of two sets, one

assuming all regressors are purely 1(1 ), and the other assuming they are all purely 1(0). 13

These bands cover all possible combinations of variables including fractionally

integrated ones into 1( 1 ) and 1(0). The null hypothesis postulating no long-run

relationship would be rejected (not rejected), if the computed F-statistic were higher (lower) than the upper (lower) band of the bound of the critical value. If the computed F-statistic falls outside this band, a conclusive decision can be made without any

knowledge of whether the underlying variables are 1( 1 ) or 1(0). On the other hand, if the

F-statistic falls within the band, information on the integration is necessary before making a decision regarding the long-run relationship.

In the second stage, if the variables in each equation are found to be cointegrated, the dynamic structure of the equations can be estimated using the ARDL procedure. 14 The dynamic structure of the ARDL (p,qj model takes the following form:

(4. 1 1 ) where XI is an endogenous variable, a is an intercept, Z/ is a vector of e xplanatory

variables, p and q are, respectively, the lag lengths of X; and Z/" and )J. is the random error term. Finally, the goodness of fit criteria and properties of the models are given in the diagnostic tests. They consist of the Durbin-Watson (DW) test for autocorrelation, normality of the residuals based on a test of skewness and kurtosis, autoregressive

conditional heteroscedasticity (ARCH), and the Ramsey RESET test for the model

specification.

1 2

See Pesaran and Pesaran, 1 997 (pp. 477-478).

\3 This method avoids the problem of serial correlation that arises in the residual-based cointegrated methods by an appropriate augmentation (Pesaran et aI., 1 996).

1 4 The orders of the lags in the ARDL model are selected using one of the four choice criteria. They are Theil's ( 1 97 1 ) R-Bar Squared criterion proposed by Pesaran and Smith ( 1 994); the Akaike Information Criterion (AlC proposed by Akaike ( 1 973); the Schwarz Bayesian Criterion (SBC) proposed by Schwarz ( 1 978); and the Hannan-Quinn Criterion (HQc) proposed by Hannan and Quinn ( 1 979). Refer to Pesaran and Pesaran ( 1 997, pp. 352-355) for the properties of the last three criteria.