variations than manufacturing. Thus, the services sector has a softening effect and cushions the economy from the drastic and harmful effects that filter from the vulnerable manufacturing sector - which are more pronounced due to Singapore’s openness. Several reasons have been put forth to explain the resilience of the services sector.
One is that service output cannot be stored and thus the services sector is spared the effects of swings in inventory investment. Elfring (1989:435) gives two other reasons for the relatively limited cyclical responsiveness of market services. First, there is more disguised unemployment in services, especially among the self-employed. Second, the flexibility of compensation systems is higher in services because of a) commission payments in many sales activities b) flexibility of the incomes of the self-employed and c) limited presence and influence of labour unions. Nusbaumer (1987:66) claims that the linkage function played by services between physical production and the market in all economies is not likely to vary in proportion with the volume of physical output, and this would serve to explain their smaller cyclical variability.
But studies show that the composition of the services sector is of importance too. For instance, producer services which are closely linked to industrial production are said to be volatile to business cycles. Should the producer services share increase, then the services sector may not be as resilient as expected. Also, if the income elasticity of services is high, then the demand for services would be affected by business cycles, thereby reducing resilience.
2.4 Causality Between the Manufacturing and Services Sectors’
Growth
Here we examine if growth in one sector causes growth in the other. The direction of causality of growth has important policy implications.
In the absence of an agricultural sector, a common notion is that an economy in its initial developmental stage would receive a larger share of its GDP from the industrial sector than from its services sector. But as it develops, the services sector becomes an
increasingly significant contributor to GDP.“ The implication here is that the growth of the industrial sector causes the services sector to grow. This is what Bhagwati (1984) termed as the ‘splintering effect’ where services grow due to increased production of manufactured goods.23 Such a trend has already been identified for Singapore by Toh and Low (1988, 1994). This is the case when shipping services, advertising, marketing and commerce thrive because of the need to sell manufactured goods locally or abroad. Also, with increasing specialisation, or technological progress allowing for standardisation of services or relatively lower transactions costs, industrial firms may choose to contract out their ‘in house’ services, thereby increasing the link from the manufacturing to the services sector.
But services sector growth could also influence growth in the manufacturing sector. Bhagwati (1984) explains this as the ‘disembodiment effect’ where goods splinter off from services due to a technical revolution in information and communication technology. For instance, R&D activities could result in improvement in manufacturing technology and thus an increase in industrial output. When banks provide low borrowing rates, there is an incentive for manufacturers to borrow and produce more. The existence of trading companies and their worldwide network can be seen to encourage greater exports of manufactured goods as producers now increasingly rely on such middlemen (who have specific knowledge) to conduct their trade for them. However, these examples seem to suggest that the causality from services to the industrial sector is likely to take place in the later stages of an economy’s development.
A causality study by Kalirajan and Kapuscinski (1993) on the experience of selected Asian countries provides some evidence for uni-directional causality running from the industrial sector to the services sector. This is the case for Philippines, Thailand, Japan, India and Malaysia which show that the pattern is similar for countries at varying levels of economic development.
The objective here is to investigate whether such causal links between the
22 It is however acknowledged that in some developing countries, labour surplus tends to be absorbed in service occupations as domestic servants and government employment to a significant amount.
23 But not all services grow from the industrial sector. For instance, public administration, law and order, education and health are quite independent of the growth of manufacturing.
manufacturing and services sectors exist in Singapore. The testable hypothesis, following Bhagwati’s ‘splintering effect’ is then formulated as:
Growth in the manufacturing sector causes growth in the services sector.24
Methodology of the Causality Tests
Drawing on Kalirajan and Shand (1992), consider the following simple structural model of Jacobs (et al., 1979).
x t = a y , + a ux t_x + a 12y,_i +w„
y , = ß x, + a 2lx,_, + a n y,_, +w2,
where x and y are two time series variables and wu and w2t are independent, serially uncorrelated random variables distributed as N(0, Gi2 ) and N(0, G22 )
respectively. The aim is to examine the causality between x and y. The reduced form for the structural system above is as follows:
xA f x \ = (k ) yj , + \ y j e, \ r - 1 \ £ 2 J where (n) = (1 - aß) 1<ctw Ka n + a a 2Xa n + a a 22^ + ß a ua 22 + ß a n > and (*1 A v c 2 y = (1-ocß)-1f 1 ( wi> ß 1J
The extent to which y influences x is described by three hypotheses:
1) H0 : a = OC12 = 0
This hypothesis is that, neither the current nor the past effect of y is transmitted to x , that is, y does not cause x .
2) H0 : a = 0
This hypothesis is that the current influence of y does not affect x . 3) H o : CX12 + OCCX22 = 0
24 This is to say that any negative growth in the manufacturing sector would cause negative growth in the service sector as well.
This is often referred to in Granger’s sense as y does not cause x. It is equivalent to testing if nn is zero. If K\2 is not equal to zero, then y causes x. However if 7112 is equal to zero, then it is impossible to conclude that y does not cause jc . This is due to
the fact that both a and a n may not be equal to zero.
As the structural model is not identified, its parameters cannot be estimated and so, a and (X12 cannot be estimated either. Thus, only the third hypothesis can be tested. For this exercise, it is appropriate for the methodology to be based on a bivariate case. Then the implicit assumption is that all variables except the rate of growth of services and the manufacturing sectors may be excluded from the analysis without giving rise to spurious causality.
For Granger’s causality, the following linear models are estimated:
Serv = f { Manu, Manu (-1), Serv (-1) } (1)
Manu = f { Serv, Serv (-1), Manu (-1) } (2)
A uni-directional causality from Manu to Serv requires that coefficient estimates of Manu or Manu (-1) is significantly different form zero in Equation (1) and that Serv and Serv (-1) in Equation (2) is not significantly different from zero.
However, the literature indicates that when different methods of testing are used to investigate the same relationship, results tend to vary. Thus, another method by Sims (1972) is applied to test the above causality relationship. Sims uses a two-sided distribution method while Granger uses a one-sided distributed-lag method.
Sims’s Forward-Backward Regression Model
Serv = f { Manu, Manu (-1), Manu(+1)} (3)
Manu = f { Serv, Serv(-l), Serv (+1) } (4)
If causality runs from Manu to Serv only, then Manu (+1) would not be significantly different from zero.
All 4 equations above were then estimated by the ordinary least-squares method with a time trend and a constant using the computer software, Shazam.
Data
Empirical testing is done using the real (measured in 1985 market prices) annual GDP growth rates of the two sectors over 1960-1994 and the data is drawn from various issues of the Singapore Yearbook of Statistics.
Testing Stationary of Series
The empirical evaluation of causality is dependent on the stationary properties of the two sectoral growth rates series. Non-stationarity of time series may contribute to spurious regressions and thus can significantly alter tests of the causal relationships between the sectors. Although growth rates are used to induce stationarity, it is still necessary to check on the autocorrelation properties of these series.
To do so, the unit root test of Dickey and Fuller (1981) is used. Consider the following ‘augmented Dickey-Fuller (ADF)’ regression:
k
Ax(t) = y + 8 t + £ x ( t - 1) + X ßi A x(f-i) + £{t)
i = i
The test statistics O i O 2 and O 3 are designed to test the following: a) Ho : C = 0
b ) H0 : Y= 8 = £ = 0 (presence of drift)
c) H0 : 8 = f = 0 (presence of deterministic trend)
with the alternative hypothesis in each case being the stationarity of the series.
To ensure that the series is uncorrelated, the lag structure was selected based on minimising the Akaike Information Criterion (AIC). It was found that the optimal number of lags identified by Shazam for both series was zero. The results of the unit root tests are reported in Table 2.4 below.
Table 2.4 : Stationarity Test Results
Test Manufacturing Services Critical Values
(5% level of significance)
<E>, -3.628 -3.708 -3.6
O 2 4.408 4.591 5.68
<i>) 6.610 6.879 7.24
For O i, the test values (in absolute terms) are above the critical value and therefore statistically significant, indicating that the null hypothesis of non-stationarity cannot be accepted. Hence, for the causality testing, the growth rates of the series can be used in their levels. However for tests O2 and O3, the conclusion is that the series is non stationary. Furthermore, as the test values forO 1 are very close to the critical values, it was decided to test for the stationarity of the first differences of the series.
This then requires the estimation of the following ADF regression: k
A2 x(t) = y , + f , Ax(/-1) + A2x ( t - i ) + e ,(0 i = i
The null hypothesis tests whether f, = 0 . The test values for the manufacturing and services sectors are - 4.217 and -5.459 respectively while the critical value for 5% level of significance is -3.6 . Thus the results indicate that both the series are stationary of order one or I (1).
The causality tests were then performed on both series when they are stationary of order zero or I (0), and I (1). As similar results were obtained in both cases, only the test results of the series used at their levels are reported.
Empirical Results of Causality Tests
A summary of the results of Equations (1) to (4) are tabulated below and the significance of the test statistics at the 5% level is indicated with an asterisk, *.
In Table 2.5, both the Granger and Sims tests show that Manu (t) is significantly different from zero and Manu (t+1) is statistically insignificant as expected. Thus, causality is seen to run from the manufacturing sector to the services sector.
In Table 2.6, both tests show that Serv (t) is significant and the Sims test shows that Serv (t+1) is not significant as would be expected. Thus, causality is seen to run from the services sector to the manufacturing sector.
Table 2.5 : Causality from Manufacturing Sector to Services Sector
Variable Granger Test Sims Test
(f-Test Statistics) (f-Test Statistics)
Manu (t) 3.417* 3.323*
Manu (t-1) 0.007 0.713
Manu (t+1) - 0.059
Serv (t-1) 1.447 -
Table 2.6 : Causality from Services Sector to Manufacturing Sector
Variable Granger Test Sims Test
(f-Test Statistics) (f-Test Statistics)
Serv (t) 3.417* 3.397*
Serv (t-1) -0.825 -0.063
Serv (t+1) - 0.749
Manu (t-1) 1.392 -
In all of the above tests, there were no problems with serial correlation or heteroscedasticity. The R2, which ranges from 0.3 to 0.4, is small but not of great concern, as the tests here are merely looking for the existence of a relationship and not for the major determinants of the dependent variable. The stability of the regression coefficients was also tested using the Chow test with a structural break in 1980. There was no evidence of such a break, and thus the above regressions hold for the period of
1960-1994.
Interpretation of Results
The above analysis shows that there is bidirectional causality between the growth of the services and manufacturing sectors, indicating the presence of Bhagwati’s ‘splintering’ as well as ‘disembodiment’ effects. More interestingly, the feedback is immediate as its influence takes place within the same period.26 This further reinforces the common notion of ‘servicised’ industrial and ‘industrialised’ service activities, but in reality the
26 However, if the feedback occurred within months, it would not be picked up as the data used was