Journal of Agn'culliiral Kconomirs - Volume 51, Number 3 - .Y@lmh 2000 - P a p 353-370

Agricultural Growth and Inter-Sectoral Linkages in a Developing Economy

N. Gemmell, T. A. Lloyd and M. Mathew (Manuscript received March 1998; Revision accepted June 2000)

D oes growth in the manufacturing sector of an economy spillover to agnculture, or do sectors share similar growth rates only when thqr share some common exogenous stimuli? The limited number of investigations of this issue, fbr cross-sections of countries, have found

some evidence in favour of spillovers, though the methodologzes used cannot readily separate correlation from causation. Adapting the Feder (1 982) model of sectoral externalities to a timeseries context, we examine how fur agn'cultural output in Malaysia has been aflected by inter-sectoral spillovers. Our results suggest that expansion of manufacturing output, though ussociated with reduced agn'cultural output i n the short-run, is associated with agn'mltural expansion over the long-run. Service output growth on the other hand seems to have been inimical to agricultural growth in both the short- and long-runs, whik causality testing supports the case fm spillovers rather than "common causes''. Evidence on sectoral productivity is consistent with neoclassical arguments suggesting that the benefits of higher productivity i n manufacturing tend to spill over to agriculture,. encouraging productivity convergence.

1. Introduction In the seminal contributions of Lewis (1954) and Fei and Ranis (1964) agricultural development assists the expansion of manufacturing. Agriculture is a source of (surplus) labour and savings required for industrialisation, a source of inputs for industrial processing (e.g., food, textiles), and a potential source of demand for manufactured products such as machinery, fertiliser and processed foods. Some of these sectoral interactions have been modelled more explicitly in recent years, for example via in terse.ctora1 spillovers or productivity differentials (see Feder, 1982, 1986; Hwa, 1989; Dowrick, 1990; and Dowrick and Gemmell, 1991), to capture what Hwa (1989, p.107) calls "the relationship of interdependence and complementarity between agriculture and industry."

A lack of sufficient time-series data for individual less developed countries (LDCs) has meant that cross-section regression techniques have dominated empirical investigations

1 Norman Gemmell is a Professor of Economics and Tim Lloyd a Lecturer of Economics, both at the Centre for Research in Economic Development and International Trade (CREDIT) in the School of Economics, University Park, University of Nottingham, NG7 2RD, UK Manna Mathew is a former research student in the School. The authors are particularly grateful for a number of useful suggestions made by one of the referees and the External Editors of the Journal: The usual disclaimer applies. ' While the predominant sectoral distinction considered has been agriculture-industry, services (Dowrick, 1990) and export/nonexport sectors (Feder, 1982) have also been analysed.

354 N. Gemmell, 7: A. Lloyd and M. Mnllrav

of these models. This approach suffers from two major drawbacks, however. Firstly, while growth rates across different sectors have typically been found to be positively correlated, it remains unclear whether this reflects spillovers from one sector to another (or multi- directional spillovers) 1 or both sectors responding to some common exogenous factor (e.g. growth in export demand or technical progress) .z Secondly, (unobserved) long-run time-series behaviour is often inferred, though controversially, from the cross section evidence. Even if this is accepted, the cross-section growth regression methodology cannot provide evidence on short-run behaviour, yet it might be expected that any positive externality effects from one sector’s expansion would be mitigated in the short- run by sectoral competition for (relatively fixed) inputs.

In this paper we report an investigation into the issue of inter-sectoral spillovers and interactions between agriculture, manufacturing and services in Malaysia - an LDC which has undergone a rapid transition in recent decades from a predominantly agrarian economy to one increasingly composed of sophisticated industrial and service activities. Adapting the Feder (1982) model to a time-series context suitable for our data set, and applying time-series techniques enables us to address both the issues of causality (at least in terms of “temporal precedence”) and the identification of long- and short-run interactions between sectors. One reason why it is important to understand these interactions is that government policies in LDCs such as Malaysia are often aimed explicitly at boosting the output of particular sectors (in addition to those aimed a t “underlying causes“ such as capital accumulation), or they implicitly favour certain sectors, for example by protecting different sectors to varying degrees. To the extent that such policies affect the outputs of the targeted sector(s), our results may therefore provide some guidance on the unintended (non-targeted) sectoral outcomes of such policy interventions.

The remainder of the paper is structured as follows. Section 2 focuses on the niodel of Feder (1982, 1986), adapting it to a time-series context which requires a model in levels rather than first differences. We also demonstrate that the model can be readily re- specified in terms of sectoral output interactions thus circumventing the requirement for suitable factor endowment data that are typically unavailable for developing countries, such as Malaysia. Following a brief review of the relevant aspects of the Malaysian economy in Section 3, Section 4 sets-out our econometric approach which uses the properties of time series data to gain insights into the dynamics of structural change. Sections 5 and 6 discuss the results that clearly demonstrate the presence of sectoral interactions in terms of both income level and labour productivity. Furthermore, these interactions are found to differ between sectors and across the long and short runs. Overall, the evidence suggests that for the Malaysian economy at least, agriculture has played a passive yet complementary role in economic development, and it is with this message that the paper concludes in Section 7.

‘ In fact, both Hwa’s (1989) and Feder’s (1982) modelling approaches only consider interactions/spillovers in one direction: from agriculture to industry (Hwa) and exports to non-exports (Feder, 1982). Interestingly, unlike Hwa, Feder (1986) predicts spillovers from industry to agriculture. ‘ Criticisms of the Feder methodology have been articulated mainly with respect to the Ram (1986) application, in which Ram replicates the Feder (1982) model but where export/non-export sectors are replaced by government/nongovernment sectors. See the Ycommentsn in Amnican Economic Rnn’nu, 79 (1989).

Agrirultural Crorirlh and In.ter-St.ctom1 Iinkages in n Developing Economy 355

2. Modelling Inter-sectoral Linkages To examine linkages between sectors of a growing economy, Feder (1982) proposed a neoclassical “sources of growth” model incorporating disequilibrium (in the form of productivity differentials) and spillovers between sectors, based on an export/non- export distinction. Most subsequent investigations of the growth effects of agriculture- industry differences have been based on this model - see Feder (1986), Hwa (1989) and Dowrick and Gemmell (1991). To investigate similar issues in Malaysia we use the Feder model as a starting point.

Feder specifies production functions for two sectors as follows:

A = F(K,, La, M) (1)

where A (M) is agricultural (manufacturing) output, K, (Li) is capital (labour) in sector i (= a, m denoting agriculture and manufacturing respectively). The term in M in (1) can be regarded as capturing externalities from manufacturing output in agriculture “since they are not reflected in market prices” (Feder, 1982, p. 61); or following Romer (1986), firms (sectors) are assumed to ignore the outputs of other firms (sectors) in their profit- maximising decisions over inputs and outputs. Lacking data on sectoral inputs (for a cross-section OF countries), Feder (1982) shows that the two sector production functions can be solved in first differences to express aggregate output growth in terms of uggregule input growth and the growth rate of the externality-originating sector. For Feder’s two- sector case and cross-country data this approach provides a neat solution to his data deficiencies.

To apply the Feder model in our context, however, we need to adapt it in four ways. First, to explain long-run behaviour using time-series data requires a model in levels rather than first differences. Second, we generalise externality effects to allow for the possibility of spillovers to, undfrom, agriculture. Third, as we show below, applying the model in levels requires data on (aggregate) capital and labour inputs rather than investment and labour growth, as in Feder’s case. Since suitable capital stock data are not available in Malaysia1 we demonstrate that the model can be solved to eliminate aggregate inputs. This considerably reduces the data requirements and overall complexity. Fourth, since a larger fraction of output in Malaysia (as in many countries) is generated in the service sector than in either manufacturing or agriculture, significant interactions may be missed if this sector is ignored. We therefore extend the Feder model to three sectors. However, since this extension is analytically straightforward (yielding additional empirical rather than analytical insights), we simpliEj the exposition by deriving the two-sector case below before introducing the service sector. The mathematics of the three sector case is Dresented in an amendix. ’ Such constraints can sometimes be overcome by summing (and depreciating) data on aggregate investment over a number of years prior to the period of investigation. Unfortunately such data are unavailable and/or unsuitable in Malaysia.

356 N. GemmeU, ‘I: A. Lloyd and M. Matheu

Though services have typically been ignored in previous models (an exception is Dowrick, 1990) a largely separate literature has recognised sectoral linkages involving the services sector (see Fuchs, 1968; Blades et al., 1974; Gemmell, 1982; Bhagwati, 1984). For example, the demand for intermediate services such as distribution and retailing from both agriculture and manufacturing are obvious and frequently observed (from input- output tables) to increase over time in LDCs. In addition, “final use“ services can be close complements (or substitutes) in demand for agricultural and manufacturing products, and the service sector often competes directly with agriculture for labour. A ~ r i O r ; , therefore, important interactions may be missed if the service sector is ignored, and our empirical analysis seems to confirm this in practice. However, to introduce the principles that underlie our Ssector model, consider the following forms of (1) and (2):

Marginal products are measured by ai and Pi, and externality effects by 7/i which allow for the possibility of two-way spill-overs. Adopting the Feder (1982, 1986) assumption of marginal productivity differences between sectors such that:

the coefficient 6 represents a measure of the efficiency of resource use in manufacturing relative to agriculture. Adding (1*) and (2”) it can be shown that:

Y = A + M = cp + a, L t PaK+ (6/1+ 6) 1% L, + P, K,) +?la M +ym A (4)

where L = La t L,,,, K = K , t K, and cp = cpa + cpm are economy-wide aggregates. Note that from (2*): L, + P, & = M - qm t y,A which, after some manipulation, yields:

. Equation ( 5 ) is the equivalent in levels of the first difference

ApicTinilturnl Growth imd Intpr-Sertornl Iinkiiges in n Dnalqbing Econorny 35 7

expression derived by Feder. The term {w} captures the effects on agriculture

of expansion in manufacturing output, for given factor endowments, K and L, and may be positive or negative. Pure (positive) externality effects (ym, ?la) have positive effects on agriculture while the effect of the productivity differential is to reduce agricultural output, for 'ya < 1. This latter effect arises because expansion of manufacturing, conditional on fixed total factor endowments, competes inputs away from agriculture.

Given data on capital stock we could proceed to estimate equation (5). In the absence of such data however, the model can be adapted to eliminate aggregate inputs. Firstly, we adopt the Feder (1982, 1986) assumption that sectoral marginal products of labour are related to average productivity in the economy as a whole such that: a, = a(Y/L). Secondly, we make a similar assumption for capital productivity: pa = P(Y/K). This allows us to substitute for ~ l , and Pa in (5) which, after some re-arranging (recall: Y = A + M), gives the following relationship between agriculture and manufacturing:

ym ) (1 - (0: t p) -- + 6 . Thus, larger values of a and/or p imply larger

endowments, a larger fraction will go to agriculture if the sector's productivity of capital and/or labour is relatively high. There'is also an additional spillover from manufacturing here to the extent that an expansion of manufacturing output, M, increases Y/L and/or Y/K This raises agricultural marginal products, a,, Pa, and thereby agricultural output. Notice that (6) cannot be regarded as a conventional reduced form relationship since both A and M are potentially endogenous.' In our empirical analysis therefore we adopt a vector autoregressive (VAR) framework which treats all variables as potentially endogenous. Since this approach allows us separately to identify both short- and long-run effects, we might expect that short-run evidence would approximate the effects of expansion in M, identified in (5), rather than (6), that is, where factor endowments are approximately constant. If, in addition, inter-sectoral spillovers take some time to come through, short-run effects of manufacturing expansion will primarily reflect the impact of resource competition, (as captured by the parameter element-1/(1+6) in (5) and (6)).

put. This reflects the fact that, out of any expansion of factor

Finally, in the Appendix, we derive three-sector equivalents of equations (5) and (6) which reveal analogous relationships between agriculture, manufacturing and services. As with the two-sector case, expansion of either manufacturing or services can have positive or negative net effects on agriculture depending on the size of sector (marginal) productivity differentials, inter-sectoral externalities and sectoral competition for resources.

~~~ ~ ' This endogeneity problem also afflicted Feder's applications, even though endowments were included, since one sector's output appeared as a right-handside variable.

358 N. Cemmell, 7: A. Lloyd rind M. Mathew

3. Economic Development in Malaysia The structure of the Malaysian economy has changed radically in the last twenty-five years to the extent that it is no longer dependent on a few primary commodities. The production base has broadened with the manufacturing sector’s share of GDP rising from around 10 per cent in 1965 to more than 25 per cent by 1990. Over the same period, though service sector GDP as a whole fluctuated in the 42-48 per cent range, there was considerable growth in “modern” services, while agriculture’s contribution to GDP declined from about one third to less than 20 per cent. Nevertheless, agriculture is widely considered to have played an important part in sustaining a real rate of economic growth of around 7-9 per cent per year with successful modernisation of the plantation sub-sectors (rubber, oil palm, and cocoa).

Like other Asian countries, rapid industrialisation has become a major objective of Malaysian development policy since the 1960s. Motivated by the desire to diversify away from a perceived over-reliance on primary production and, more generally to modernise the economy, exports of manufactures have been strongly encouraged. In recent years the share of manufacturing exports has doubled and now accounts for around 70 per cent of all Malaysian exports. Growth has been especially strong in the non-resource based industries of electronic<. electrical machinery and textiles.

Table 1 Seclor I965 1975 1985 1990’ 19952

Percentage Contributions to Malaysian GDP (1980 Prices)

Agriciilture 31.5 27.7 20.8 18.7 13.0

Man iifac triring 10.4 16.4 19.7 27.0 33.0

Services 44.6 47.5 44.2 42.2 44.0

Source: Yap and Nakamura (1990). I Government of Malaysia (1991).

World Bank (1997), current price share.

Since a consistent set of data on the main sectoral GDP aggregates for Malaysia is available for 1965-91 this forms the sample period for our investigation of sectoral GDP linkages. Data on employment are only available from 1970 and sectoral definitions do not exactly match those for output, so our productivity comparisons are necessarily more limited and should be interpreted cautiously. Sectoral GDP (at 1980 prices) and productivity profiles are given in Figures 1 and 2 (see Sections 5 and 6) respectively and indicate generally increasing GDP and productivity. After a long period of rapid growth, the mid-1980s represent something of an interruption in these trends however, with a slight decline in manufacturing and service GDP in 1985, and a large fall in service GDP (and productivity) in 1986. The reasons for this recession (unprecedented in Malaysia’s recent development) and its particular severity for services are discussed in Ariff (1991). The major impact on services arose from a unique retrenchment in government services and major depreciation of the currency in 1986 that dramatically raised the relative price of non-tradables (mainly services). For present purposes, its significance lies in the need to model a temporary shock to GDP and productivity trends in services (see Section 5).

Agrinrllurnl Growth nnd Inter-Sectcml Linkages in a Developing Economy 359

4. Econometric Methodology As noted earlier, to estimate a three sector version of equation (6) we use a vector autoregressive (VAR) framework, which allows the data to determine the precise niodel specification and treats all variables as potentially endogenous. Given the familiarity of these methods we merely offer a sketch of the relevant aspects here.’ Consider the VAR(k), model:

where t = 1, ..., T; xt is an (n x 1) vector of endogenous variables and p is an (n x 1) vector of constants and E,, an (n x 1) vector of independently distributed disturbances of zero mean and nondiagonal.variance-covariance matrix 0, i.e. E, - n.i.d. (0,Q).

Following Johansen (1988), if the variables cointegrate then (7) has an equilibrium correction representation given by

comprising an (n x r) matrix of cointegrating vectors, B, and an (n x r) matrix of “equilibrium correction” coefficients, A. Empirically, the number of long-run relationships (denoted by r) , is determined using the trace and maximal eigmvalue test statistics. Should r = 1 there cxists a single cointegrating vector: the elements of B quantify this long-run equilibrium relationship and the elements of A measure the rate at which disequilibrium is corrected in the system. If the short-run impact of shocks on Ax, differs from the long-run response then this is captured by the coefficients of Ti. Thus, the short and long run responses to shocks are allowed to differ.

In the current context, we may consider “equilibrium relationships” as representing enduring inter-sectoral linkages that bind sectors together in the process of economic development as the growth of one sector reflects, inter a h , the size and state of others with which it interacts. To the extent that resource competition, productivity differentials or spillovers between sectors induce long-lasting (linear) effects, those long-run relationships should be evident through the coefficients of B.

A VAR model also facilitates investigation of the related concepts of exogeneity and temporal precedence, or more commonly, Grangercausality (Granger 1969). Whereas single equation methods force exogeneity of the explanatory variables by assumption, a system-based approach allows these assumptions to be tested empirically, via parameter restrictions. Johansen’s (1992) test for “weak exogeneity” (based on the notion that variables that do not respond to disequilibrium in the system of which they are a part, may be considered (weakly) exogenous to that system), tests the significance of specific ’ Those seeking an induction are referred to Harris (1995) or Enders (1995).

360 N. CemnreU, 7: A. Lloyd and M. Mathew

elements in A of (8). In the case where r = 1 , weak exogeneity of (n - 1) variables reduces the complexity of the modelling exercise and legitimises the use of single equation methods. Furthermore, for a VAR( 1 ) model in which components cointegrate (the case below) weak exogeneity implies Granger noncausality.1 In the current context, these results signal whether the adjustment mechanisms - or transmission channels - between sectors are unidirectional or multidirectional.

In sum, vector autoregressive methods offer a natural framework for the study of structural change - an inherently inter-temporal phenomenon - allowing previously untested aspects of the process to be addressed systematically. Caveats are nevertheless warranted, not least since VAR methods have been the subject of critical assessment on methodological grounds (inter alia Darnel1 and Evans, 1990). Of particular relevance here is that the asymptotic theory upon which inference is based rests uneasily with the small samples at our disposal and the inherent over-parameterisation of system-based approaches. We report finite sample corrections of the test statistics below, but the exact finite sample distributions are unknown and thus adjustmen'ts for sample size are necessarily approximations to the true critical values (Cheung and Lai, 1993). We also refer to results from single equation estimation for corroboration of the system-based results where inference is not clear-cut.

5. Testing for Sectoral Interactions

We now proceed to analyse the data on Malaysian sectoral GDP using the abbreviations: at, m, and st to denote the logarithms of sectoral GDP (1980 prices) in agriculture, manufacturing and services respectively. D, denotes an intercept dummy ( D, = 1 when t = 1986, zero otherwise) to accommodate policy change in 1986.'

To test for the order of integration of the series and the existence of cointegration a VAR model is estimated. Mindful of the small sample we have at our disposal, a V4R(2) model of equation (8) is specified comprising the variables a,, m,, st and D,. Following estimation, system-based tests for residual autocorrelation, normality and heteroscedasticity do not uncover departures from the stated assumptions at conventional significanceAlevels, although the system does appear to be over parameterised testing that 112 is a null matrix is easily accepted (F(9,36)=10.8 (p value =

0.40)). This reduction delivers the VAR( 1) system with "white noise" residuals and an R* of 0.99 that is depicted in Figure 1, and is the model to which the tests for cointegration are applied (reported in Appendix Table A l ) . Aware that "over-rejection" (of the null hypothesis of no cointegration) can occur when asymptotic critical values are applied to small samples (Reimers 1992), we also report the finite sample critical values proposed by Cheung and Lai (1993) in Table A l . The test statistics clearly reject the null using the asymptotic critical values, but reject marginally below 10 per cent when finite sample

' In general, additional restrictions on the coefficients in r, are also required to demonstrate Granger non- causality (see Johansen, 1992; Mosconi and Giannini, 1992) but are clearly redundant in the VAR(1) case. For obvious reasons Hall and Milne (1994) have called this form of Grangercausality %eak causality". ' The statistical analysis is conducted using PCFiml9.0software (Hendry and Doornik, 1996). In the interest of brevity only key results are reported. Ail data and detailed results have been made available to the referees and may be obtained from the authors upon request.

Agrirullural Graiulh nnd Inter-Serlmal Linkages in a Dmclcj&zg Economy 361

critical values are applied. In view of the strong evidence in favour of cointegration in (10) below and additional results obtained using single equation methods, we infer the existence of a single cointegrating vector between sectoral GDPs in Malaysia.' With three variables in this system and one cointegrating vector there are (n - r =) 2 unit roots, implying that all the variables are I(1) processes, as is confirmed by conventional univariate tests of non-stationarity.

Normalising on at suggests a long-run relationship between sectoral GDPs of the form2

a = 0.67m - 0.47s (9)

implying, c e t a ' s pan'bus, a 1 per cent increase in manufacturing (services) GDP leads in the long-run to a 0.67 per cent rise (0.47 per cent fall) in agricultural GDP. While this normalisation is arbitrary, as writtcn it implies the exogeneity of manufacturing and services GDPs to agricultural GDP. The small magnitude of the coefficients in A corresponding to the manufacturing and services equations (-0.1 and -0.17 respectively) lend weight to the normalisation in (9). More formal evidence is given by Johansen's (1992) test, that yields a x 2 ( 2 ) likelihood ratio test statistic of 2.06, (p value 0.36), indicating the weak exogeneity of manufacturing and services in the equilibrium relation.

Figure 1 The VAR( 1 ) Model of Sectoral CDP (Log) Actual and Fitted Values Standardised Residuals

' All single equation tests based on an autoregressive distributed lag model (ADL) signal cointegrarion at the 5 ,yr cent significance level. Esumates from the single equation ADL model are virtually identical (i.e., a = 0.63m - 0.41s).

362 N. Gmmell, 7: A. Lloyd and M. Mnlhew

This result implies that in the long run it is only agriculture that adjusts to sectoral disequilibrium within the economy and (by virtue of the VAR( 1) specification) indicates that agriculture is not a Grangercause of growth in the manufacturing or service sectors. That is, shocks in manufacturing and services GDPs appear to spill over to agriculture but not vice versa In the Malaysian case therefore, agriculture would appear to assume a passive role in the development process, reacting to, or benefiting from, changes elsewhere in the economy, but changes in agricultural output do not appear to be important for growth elsewhere in the economy.

From (9) the long-run linkages to agriculture appear to give rise to quite large, positive effects with respect to manufacturing, but somewhat smaller, negative effects from services. Unlike manufacturing, service expansion would appear to be detrimental to agricultural growth in the long-run. Since, according to our model in Section 2, these parameter estimates capture the combined impacts of externalities, productivity differentials and competition for resources, it would seem that this final effect dominates between services and agriculture. That is, service sector growth is sustained by competing inputs such as labour away from agriculture, and equation (9) suggests such negative effects from services persist into the long-run.

The exogeneity test results imply that single equation methods are efficient and allow Amt = m, - mt:l and As, = st - s , ~ to enter as explanatory variables in the final equilibrium correction representation given below (t-ratios beneath parameter estimates) :

Aat = 2.78 - 0.39Amt - 0.72Ast - 0.61 (a - 0.67m + 0.47s),, (4.34X-1.86) (-5.69) (-4.40)

- R'r = 0.64

Diagnostic Testing (p - value)

Autocorrelation: F(1.21) = 1.3 (0.72)

Normality: x2 (2) = 3.33 (0.19) Heteroscedasticity: F(6.15) = 1.26 (0.33)

F(3.21) = 12.79 (0.0000) DW = 1.84

Reset test: F(1.21) = 0.03 (0.86)

The coefficients of (10) have the following interpretation. The underlying rate of growth in agriculture is estimated to be 2.8 per cent per year, a little over half the actual average growth rate, indicating a substantial degree of inter-sectoral interdependence. At times when the steady-state relationship between sectoral GDP does not hold, disequilibrium feeds back into agricultural growth via the equilibrium correction coefficient, at an estimated annual rate of 61 per cent, presumably through resource flows between sectors. In other words, sectoral disequilibrium does not persist for long, since the agricultural sector reacts rapidly, making half of any necessary adjustments within ten months. Given the rapid rate of economic growth of the Malaysian economy and labour market flexibility, this estimated rate of adjustment is quite plausible.

The coefficients on Amt and Ast in (10) represent the short-run (or impact) semi- elasticities and imply, ceterispan'bus, that a 1 percentage point increase in the growth rate in either the manufacturing or service sector retards growth in agriculture by around 0.4

Agn'cullural Crniulh and Inter-Sectoral Linkages in n Deuelqing Emnomy 363

and 0.7 percentage points respectively. Whilst our methodology does not permit us to idenlify the sources of these dynamic responses, as we argued in Section 2 they are suggestive of short-run sectoral competition for resources (Ariff 1991, Mathew 1996). The magnitude of the point estimales imply that agricultural GDP is more sensitive to short-run expansion of services than to manufacturing, possibly reflecting the respective degrees of resource substitutability between sectors. Certainly, i t is a reasonable expectation that short-run substitution of mobile factor inputs (primarily labour) is more likely between services and agriculture than between manufacturing and agriculture.

The results are also informative regarding the relative importance of changes in manufacturing and services on the size of the agricultural sector over the longer term. A decomposition of the predicted growth in agriculture indicates that manufacturing growth is by far the most influential source of growth in agriculture. Substituting the average annual rates of growth in manufacturing and services of 12.1 per cent and 5.4 per cent respectively into (9), the predicted annual growth rate of agriculture is 5.6 per cent (= 0.67(12.1 per cent) - 0.47(5.4 per cent)). Thus, the estimates suggest that over the sample period the positive effects of growth in the manufacturing sector have been more than three times as great as the retardation of agriculture resulting from the expansion of the service sector. The demand-related impact on agriculture of growth in manufacturing-related incomes is one possible source of this dominance. In addition, by their very nature, spillover effects take time to percolate through the economy and thus play a more significant role in the longer run. In so far as production and managerial technology originates in the manufacturing sector (e.g., via the operation of MNCs and indigenous capital intensive industries) one may expect a the agricultural sector to be more responsive to spillovers from manufacturing than from services. Note also that we do not find a significant reverse effect - exogenous expansion of agriculture does not appear to affect the other two sectors' GDPs. This Granger non-causality is consistent with labour market evidence in Malaysia that, while modern sectors can attract labour out of plantation agriculture, significant reverse flows are rare (Sulaiman 1992). When demand for agricultural commodities increases, labour expansion generally takes the form of low-skill f0reg-n migrant workers.

Whatever the underlying explanations, these results would seem to suggest that, in terms of long-run GDPs, expansion of services has been inimical to agricultural development while manufacturing expansion has enhanced agricultural growth. Furthermore, the positive effects of manufacturing growth outstrip the negative effects of services in the long run. However, in the short m n , expansion of either services or manufacturing GDP is at the expense of agriculture, as would be expected when different sectors have to compete for relatively fixed factor supplies.

6. Testing for Productivity Interactions

As we saw in Section 2, the Feder model which we adapted to examine sectoral output levels, embodies particular assumptions regarding sectoral marginal productivity and average productivity in the economy as a whole. Its implications for sector average

364 N. Gmmell, 7: A . Lloyd and M. M a h w

(labour) productivity (pa = A/La, pm = M/Lm) are readily seen by dividing equation (6) by La to give:

Differentiation of (1 1) with respect to pm reveals that:

Since, by assumption (and in our data), marginal productivity is higher in manufacturing (6 > 0), then (aem/apm) > 0; that is, increases in manufacturing productivity will be associated with a relative expansion of labour in manufacturing. Therefore in (12), dpa/apm > 0 as km> 0 so that the sign of the relationship observed between manufacturing and agricultural outputs should be mirrored in the average productivity relationship. For the three-sector case i t can be shown that a similar relationship to (12) exists for services,’ that is:

For services however, where (in our data-set) the figure for average labour productivity is intermediate between those of the other two sectors, it is not necessarily the case that (ae,/ap,) > 0. Increases in manufacturing and service productivity, ceteris paribus, will simultaneously draw labour into services from agriculture and out of services to manufacturing. Thus the sign on the productivity relationship between services and agriculture, [{,e,t, may differ from the equivalent output relationship.*

Bearing in mind the caveats regarding labour input data and, pan passu, labour productivity, we can examine whether the data support these predicted productivity relationships. In an analogous manner to that above, we test for the existence of ’ In the three-sector case of course, the terms c,, and 5, are taken from Appendix equation (A10) rather than (6) above.

Alternative models can yield different predictions of course. For example Matsuyama (1992), using a two- sector (agriculture-manufacturing) model of endogenous growth in the form of learning-bydoing in the manufacturing sector, shows that increases in productivity (and output) in manufacturipg can be associated with both lower levels and slower growth of productivity in agriculture in an open economy.

Ap'cicllurnl Crorulh and hter-Sectornl Linknges in a Deoel@i.ng Economy 365

Figure 2 The VAR(1) Model of Sectonl Labour Productivity (Logs) Actual and fitted values Standurdised Residuals

relationships between sectoral labour productivity using a VAR( 2) model that comprises the 1(1) series apt, mp,, spt (denoting labour productivity at 1980 prices in the three sectors) and D,. As before, system reduction is confidently upheld IF(9, 24) = 0.91 (p value = 0.52) 1 and tests on the residuals of the VAR( 1) system indicate model adequacy. Cointegration tests are reported in Appendix Table A2 and the data, model and residuals are shown in Figure 2.

As with the GDP results, the cointegration tests using the VAR are not decisive but are suggestive of the presence of a unique cointegrating vector, which (having normalised on the coefficient of ap,) is estimated as'

ap = 1.51mp + 0.45s~ (13)

From (13) it is apparent that improvements in labour productivities in manufacturing and services lead, in the long run, to higher productivity in the agricultural sector confirming the predictions based on (12) and (12'). In particular, manufacturing output and productivity increases both have positive effects on agricultural output and productivity respectively. Though service output expansion appears to retard agricultural output (equation (9 ) ) , service productivity growth enhances agricultural productivity. ' Single equation methods indicate the presence of a cointegrating relationship at the 1 per cent significance level and estimate the long-run relationship as ap = 1.45rnp t 0 . 3 4 ~ ~ .

366 N. Gmmell, 'I: A, Lloyd and M. Mirlhew

The long-run response to changes in service productivity is relatively inelastic (0.45) however, while the manufacturing elasticity (1.51) implies that in the long term the agricultural sector is highly responsive to productivity growth in that sector. These results imply, ceta's paribus, a tendency towards convergence of manufacturing and agricultural productivity levels over time.

'Testing for the weak exogeneity of mp, and spt in this relation yields a ~ ~ ( 2 ) test statistic of 2.26 (p value = 0.32) supporting the conditioning inherent in (13). Moreover, given that weak exogeneity in the VAR( 1) case implies Granger causality, this result suggests that labour productivity in agriculture is not a Granger-cause of labour productivities elsewhere in the economy but that labour productivities in manufacturing and services do Granger-cause productivity growth in the agricultural sector.

The final error correction model of productivity growth in agriculture is estimated as,

Aapt = 0.84 - 0.24Ampt - 0.58Aspt - 0.66(ap - 1.51mp - 0.45sp),l (5.57X-0.97) (-4.08) (-5.50)

- R2 = 0.75

Diagnostic Testing (p -value)

Autocorrelation: F(1,16) = 1.1 (0.32)

Normality: ~2 (2) = 1.25 (0.53) Heteroscedasticity: F(6,lO) = 1.07 (0.44)

F(3,17) = 17.75 (0.0000) DW = 2.15

Reset test: F(1,16) = 1.51 (0.24)

I t is apparent from (14) that the underlying rate of agricultural productivity growth is less than 1 per cent per year during the sample period, considerably below the average rate over the sample of 4.6 per cent, and underlines the important linkages in labour productivity growth. In the short term, increases in labour productivity in manufacturing and services have a negative impact on agricultural productivity, although this is statistically insignificant for manufacturing.' These negative short-run relationship(s) are again unsurprising, and are suggestive of expanding sectors being able to attract the better quality or more productive inputs, so that a short-run productivity decline in the sectors from which they are attracted (in this case, agriculture) could also be expected. The dominant short-run effect is thus one of sectoral competition, and in a manner analogous to the GDP results, it is the service sector that most keenly appears to compete with agriculture. These results may therefore lend support to the commonly held view, that for much of the agricultural labour force, it is the service sector that represents the most likely alternative to agricultural employment.

Finally, ( 14) suggests that short-run disequilibrium in labour productivities across sectors is corrected relatively rapidly by appropriate (labour) resource flows, in that about two thirds of disequilibrium is corrected per year. In sum, the productivity result., are similar to those obtained for sectoral GDP in that the inimical short-run effects of non-agricultural growth on the agricultural sector are overturned by the complementary long-run relationships. ' Estimates and standard errors of remaining parameters are qualitatively unchanged when Amp, is omitted.

Agrinillurul Growth and hlm-Sectwral Linkages in n Develojing Econorny 367

7. Conclusions In this paper we have adapted Feder’s (1982) model of structural change to a time-series con text and applied it to Malaysia. Allowing for persisting productivity differences hetwcen sectors and inter-sectoral spillovers in a three-sector model of agriculture, manufacturing and services, we have shown that, even in the absence of suitable input data, it is possible to explore Feder-type linkages between sectoral GDP and productivity in both the short- and long-run.

The wider message that our results convey is the potaliuf for sectoral complementarity during the process of agricultural transformation over the long run, rather than the more usual story of resource competition which dominates the short run. In the Malaysian case it would seem that policies which stimulate manufacturing growth will beneficially affect agriculture; policies aimed at agriculture will have negligible impact on other sectors; while policies which boost services are likely to have negative knock-on effects on agriculture (at least in terms of GDP).

In light of this, the promotion of manufacturing as part of Malaysia’s development strategy may have imparted a positive externality to the agricultural sector in the long run, despite adverse short-run effects. While our approach does not separately identify the “underlying causes” of these interactions (externalities, productivity differentials, resource competition), we have argued that the signs and magnitudes which we identify are plausible given known characteristics of the Malaysian economy. For example, we would expect short-run effects to approximate a “fixed factor” context, and our evidence on the impact of non-agricultural expansion on agriculture is consistent with this, revealing the predicted effects of resource competition. On sectoral productivity, results suggest that increases in manufacturing and services both impact positively on productivity in agriculture in the long run. These results are consistent with the predictions of a Feder-type model suggesting, inter ulia, that the spillover of productivity gains in manufacturing to agriculture encourages agricultural growth and convergent tendencies in sectoral productivity levels.

References

Ariff, M. ( 1991). The rMIllnysirrn f:’conomy: P w i j c Conneclions. South-East Asian Social Science Monographs,

Bhagwati, J. N. (1984). Splintering and Disembodiment of Services and Developing Nations. Wwkl Emnonq, 7,

Blades. D., Johnson, D. D. and Marczewski, W. (1974). .SmicrActiuifie.s in Dtvdoling Cmtnlries. Paris: OECD.

Cheung, Y-W. and Lai, K S. (1993). Finite-Sample Sizes ofJohansen‘s Likelihood Ratio Tests for Cointegration,

Darnell, A. C. and Evans. J. L. (1990). The Limits o/Economctrics. Aldershot: Edward Elgar.

Dowick, S .J. (1990). Sectoral Change, Catching Up and Slowing Down: OECD Post-war Econoniic Growth

Dowrick, S. J. and Cemmell, N. (1991). Industrialization, Catching-Up and Economic Growth: A Comparative

Enders, W. (1995). Applied Ecotlomehic Time Snies. Wiley Series in Probability and Mathematical Statistics,

Feder, G. (1982). On Exports and Economic Growth, Journal oJDeuelopmmi Economics, 12, 59-73.

Oxford University Press: Singapore.

133-143.

Oxford Bulkfin ojEconomics and Slnlistics, 55, 313-328.

Revisited, Economics I ~ t e r s , 31, 331-335.

Study across the World’s Capitalist Economies, Economic Jburnnl, 101, 268275.

Chapters 5 and 6, New York: Wiley.

368 N. Gemmell, 7: A. Lloyd and M. M a t h

Feder, G. (1986). Growth in Semi-Industrial Countries: A Statistical Analysis, in Chenery, H. B., Robinson, S.

Fei, J. C. H. and Ranis, G. (1964). DLvelupmenf o j a Labor Surplus Economy. Theory and Policy. Homewood.

Fuchs, V. (1968). The S m ’ c e Economy. New York NBER.

Cemmell, N. (1982). Economic Development and Structural Change: The Role of the Service Sector,jounial

Government of Malaysia (1991). Economic Report 91/92 Ministry of Finance: Singapore. Granger, C. W. J. (1969). Investigating Causal Relationships by Econometric Models and Cross Spectral

Hall, S. G. and Milne, A. (1994). The Relevance of P-Star Analysis to UK Monetary Policy, Economic Jotlrnui 104,

Harris, R. (1995). Using Cointegrafion Analysis in Econometric Modelling. Chapter 5 and 6, London: Prentice Hall.

Hendry, D. F. and Doornik. J. A. (1996). Empirical Economefnc Modelling using PCCiue 9.0 fm Windows, London:

Hwa, E.C. (1989). The Contribution of Agriculture to Economic Growth: Some Empirical Evidence, in Williamson, J. and Pancharnurkhi, V. R. (eds.), The BaIance Between Industry and Agn’alfure in Economic ~eue~oprnrnf . Volume 2, Sector Proportions. New York The World Bank.

Johansen, S. ( 1 988). Statistical Analysis of Cointegration Vectors, Jmirntil of Economic Dynumics nnd Confrol,

Johansen, S. (1992). Testing Weak Exogeneity and the Order of Cointegration in UK Money Demand, Jmirnal

Mathew, M. (1 996). A n Inriestigntion of the Strucfurnl Changes in theEconomic Drvdo,tmmt on Mnluysin, Unpublishrd

Lewis, W. A. (1954). Economic Development with Unlimited Supplies of Labour, Manchesler School, 22, 139-191. Matsiiyaina, K. (1992). Agricultural Productivity, Comparative Advantage and Economic Growth, JOIIRZU~ o/

Economic Throry, 58, 3 17-334. Mosconi, R. and Ciannini, C. (1992). NonCausahly in Cointegrated Systems: Representation, Estimation and

Testing, Oxfmd Bulletin of~conomics and Statistics, 54(3), 39941 7. Osterwald-Lenum, M. (1992). A Note With Fractiles of the Asymptotic Distribution of the Maximum

Likelihood Cointegration Rank Test Statistics, 0xjOt.d Bullefin o/Econaics and Stntisfics, 54, 461 4 7 2 . Rani, R. (1986). Government Size and Economic Growth: A New Framework and Some Evidence from Cross-

Section and Time-Series Data, Amm’cnn Economic RcviEw, 76, 191 -203. Reimers, H.-E. (1992). Comparisons of Tests for Multivariate Cointegration, Sfnfisfical Papers. 33, 335359. Romer, P. (1986). Increasing Returns and Long Run Growth, Journal of Political Economy, 94, 1002-1037. Sulaiman, M. (1992). Thr Plnntafion Industry in the Year 2020. Challenges nnd Responses, Institute of Strategic and

World Bank (1997). Wmid Deuelopmeni Report. New York: The World Bank. Yap, J. T. and Nakamura Y. C. (eds.) (1990). ASEAN Link: An Econometric Sfudy. Longman: Singapore.

and Syrquin, M.. Indtisfrinlisnfion and Growrh. New York: Oxford University Press.

Illinois: Irwin.

of Deuelopmnt Sfudies, 19, 37-66.

Methods, Economefnca, 37,424-438.

597-605.

International Thoinson Business Press.

12(2/3), 231-254.

ofPolicy Modelling, 14, 313-334.

Ph.D. Thesis, University of Nottingham.

International Studies, Kuala Lumpur, Malaysia.

APPENDIX A Three-Sector Fedtr Model’

Let A, M and S represent output in agriculture, manufacturing and services respectively; L and K represent inputs of labour and capital. Sector production functions are given by:’

’ Constant terms are excluded for simplicity in this three sector derivation.

where the Iota1 stocks of labour and capital are L = L, t L,,, t I.,. K = IC, t K,,, t $, respectively. Aggregating (AI)-(A3), total output is,

Y = A t M t S (A4)

= a, L., t p, K, + y:M t y; S t a,, L,, t P,,, y,, t y;, A t f,, S t OL L, t P, K, t f'M + KIA

so that,

a, = a', ( 1 + 6 , ) and P, = P, (1 t si) i = m,s

Su1,stitiiting (A5) into (A4) and noting from (A2) and (A3) that,

it can be shown after some manipulation that:

To eliminate factor inputs from equation (A8) assume that the marginal productivity of labour and capital in ;Igriculture is proportional to average productivity in the economy as a whole, such that,

Substituting (A9) into (A8) further manipulation yields:

370 N. G a m l l , 7: A. Lloyd and M. Malheru

which shows that A = f(M, S) and can be seen to take a similar form to the two-sector case in equation (6). This is the form that we use in our empirical analysis. (A detailed derivation of the above results is available from the authors upon request).

Dpjinitions and Diita Sources

Definitions and sources of the data are as follows:

GDP (at factor cost) by sector, at constant 1980 prices: World Bank, World ?izblr?s; UN Nalional Accounf S1fiListir.x Yearl~ookr (various issues) ; Asian Development Bank. K q Indicators of Deueloting Asian nnd P m j c Counfries.

Lilbotir productivity is defined as GDP divided by employment in the relevant sectors. Employment data are froin: Ministry of Finance, Malaysia. Economic Rqbort (various issues).

Sector definitions are the familiar ISIC classifications: Agriculture includes Forestry and Fishing. Services iricludes Commerce; Transport, Storage and Communications; Public Utilities; Government Services; and Other Services (Community, Social and Personal Services).

Table A1 Cointegration Tests in the GDP Model (Asymptotic (m) imd Finite Sample (T) 10 Per Cent Critical Values) 26 ~)6.~,.,swnlion~ (1 966-1 991) Ei~enwrlues: 0.53 0.28 0.03

HI) HI 7i-nre m 7' klJ< m 7'

r = 0 r 2 1 29.1 26.8 29.9 19.7 18.6 20.8

r 5 1 r 2 2 9.4 13.3 14.8 8.7 12.1 13.5

r < 2 r = 3 0.8 2.7 3.0 0.8 2.7 3.0

Notes: Critical wilues are those calculated by Osterwald-Lenuin (1992) and Cheung and Lai (1993)

Table A2 Cointegration Tests in the Productivity Model (Asymptotic (-) and Finite Sample (T) 10 Per Cent Critical Values) 21 OBservrrlions (I 971-1 991) Eigtnwnlues: 0.68 0.22 0.03

HI) HI Trace m 7' m: m 7'

r = 0 r 2 1 28.9 26.8 30.8 22.8 18.6 21.4

r 5 1 r 2 2 6.0 13.3 15.3 4.7 12.1 13.9

r S 2 r = 3 1.3 2.7 3.1 1.3 2.7 3.1

Notes: Critical values are those calculated by Osterwald-Lenum (1992) and Cheung and Lai (1993).