Defence Spending and Growth in Cyprus: A Causal Analysis

Abstract

The causal relationship between economic growth and defence spending has attracted considerable attention and has been the subject of many empirical studies. Hoping to contribute to the existing pool of literature, this paper examines the relationship between military expenditure and growth in the case of Cyprus, a small island economy, for the period 1964–99. The findings reported suggest the presence of bi‐directional instantaneous causality between defence spending and economic growth.

Keywords:

• Cyprus
• Military expenditure
• Causality
• Cointegration
• Economic growth
• Structural breaks
• JEL Code: H56

INTRODUCTION

Following Benoit's (1973, 1978) seminal contribution, the relationship between economic growth and defence spending has been the subject of extensive theoretical and empirical work (Gleditch et al., 1996; Brauer and Dunne, 2002). On the basis of the generated evidence one may tentatively suggest that this relationship cannot be generalized across countries and over time.

Military spending can affect the economy through various channels. On the one hand, such expenditure may stimulate growth through Keynesian‐type aggregate demand stimulation. Increased demand induced by higher military spending leads to increased utilization of capital stock and higher employment. Increased capital stock utilization may lead to increases in the profit rate, which in turn may induce higher investment, thus generating short‐run multiplier effects and higher growth rates. Furthermore, as Deger (1986) suggests, economic growth may also be stimulated through spin‐off effects such as the creation of a socio‐economic structure conducive to growth. On the other hand, however, such spending has been shown to have growth‐retarding effects, mainly through investment crowding‐out, inflationary pressures and the reduction of available public funds for spending and investment in other, potentially more productive and growth inducing, areas. All these channels through which military spending can influence – promote or retard – growth assume that such expenditures are causally prior to economic growth.

However, as Joerding (1986) notes, economic growth may be causally prior to defence spending. Thus, although military expenditures may affect growth through the mechanisms mentioned earlier, it is also plausible that economic growth may be causally prior to military spending. For example, a country with high growth rates may wish to strengthen its external as well as internal security by increased defence spending (Dakurah et al., 2001). Furthermore, it is equally plausible that countries with high growth rates may divert resources from defence to other more productive uses.

On the basis of the forgoing discussion there are four possible outcomes when it comes to the causal ordering between growth and military spending: bi‐directional causality between the two time‐series, unidirectional causality from growth to defence expenditure or vice versa and the absence of any causal relationship.

In the context of the preceding brief discussion of the issues involved, the causal relationship between economic growth and military spending has been the subject of extensive empirical work such as, for example, Dakurah et al. (2001), Castille et al. (2001), Dunne et al. (2001), Madden and Haslehurst (1995), Kusi (1994), Kollias and Makrydakis (1996, 2000), Nadir (1993), Heo (1998) Chowdhury (1991), LaCivita and Frederiksen (1991) and Joerding (1986). A survey of this literature reveals little consensus on either the existence of such a relationship or, when it exists, its nature and direction. Unidirectional (from military expenditure to growth, or from growth to military expenditure), bi‐directional and no‐causality have been reported. On the basis of the generated evidence and its lack of consistency, one may reach the conclusion that this relationship cannot be generalized across countries and over time since, among other things, it depends on the level of socio‐economic development of the country (or countries) involved, the sample period as well as the methodology employed.

Hoping to contribute to the existing literature, this paper addresses the issue of the presence and direction of causality between military expenditure and economic growth in the case of Cyprus during 1964–99. Cyprus, a small island state and one of the ten new European Union members in its recent enlargement, gained independence from British colonial rule in 1960. In its short history as an independent state it has known inter‐communal strife, invasion, war and occupation (Kliot and Mansfield, 1997). Despite this comparatively turbulent history, it has allocated a comparatively small proportion of its national income to defence with the defence burden (i.e. military expenditure as a share of GDP) averaging around 2.5% during 1964–99 but with significant variations in different sub‐periods reflecting changes in the external security environment as well as in defence policy (Kollias, 2001). Thus, in the pre‐1974 period, defence spending as a share of GDP averaged 1.7% during 1964–74, 1.8% during 1975–86 and 3.9% during 1987–99 reflecting a major change in defence policy with the concomitant procurement of imported military hardware. Based on ACDA estimations, arms imports as a share of total imports rose from an average of 0.98% during 1975–86 to 3.35% during 1987–97. At the same time, despite the fact that the 1974 Turkish invasion caused a 16.9% and 19% fall in GDP in 1974 and 1975, respectively, and around 38% of Cypriot territory was since then under occupation, the Cyprus economy has exhibited some comparatively high and fairly steady growth rates, averaging around 5.5% during the period under consideration here (Georgakopoulos, 1999; Kammas, 1992).

Given the security and military situation on the island, as well as its economic performance, Cyprus presents an interesting and, in many respects, a unique case to examine the causality issue between military spending and growth. This is done in the next section of the paper using cointegration and causality analysis.

EMPIRICAL RESULTS FROM COINTEGRATION AND CAUSALITY TESTS

In line with previous work on the causal ordering between growth and military expenditure (see for instance Dakurah et al., 2001; Dunne and Vougas, 1999; Dunne et al., 2001), the preliminary step in our analysis is concerned with establishing the degree of integration of the variables involved (i.e. military expenditure as share of GDP, mexp_t , and GDP growth rates, g_t , that are, respectively, drawn from SIPRI Yearbooks and IMF's International Financial Statistics). To this effect, the variables are tested for the existence of a unit root in their level and first differences.

In Table 1, the findings from the application of three testing procedures, namely the standard ADF unit root test, the Kwiatkowski et al. (1992) stationarity test (KPSS) and the Zivot and Andrews (1992) sequential unit root test are reported. The application of the latter test reveals, as one would expect from a graphical inspection of the two time series in Figure 1, that both g_t and mexp_t undergo shifts in their mean (and possibly a change in trend for the case of g_t ) around the time of the invasion of Cyprus by Turkey in 1974. The consequences of the invasion were devastating both in human and material terms. The economy received a heavy blow and, as mentioned earlier, GDP fell in real terms by 16.9% in 1974 and 19% in 1975, to recover with growth rates of 18% and 15.8% in 1976 and 1977, respectively. This structural break in g_t is picked up by the Zivot–Andrews test in Table 1. It assumes its minimum value t (λ _imf = 0.33) of −9.656 in 1975 which clearly exceeds the −5.08 asymptotic critical value reported by Zivot and Andrews (1992). Hence, g_t can be described as a trend stationary process suffering a shift in mean and change in trend as a result of the Turkish invasion. A similar break is found in the case of mexp_t , occurring in 1986 with the Zivot–Andrews t _imf (λ = 0.64) statistic amounting to −5.052 and, consequently, is unequivocally significant at the 5% significance level (the asymptotic critical value is −4.8). This shift in the variable concerned is the result of a significant change around 1986 in defence policy aimed at enhancing the military capability of Cyprus so as to present a more credible deterrence and defence vis‐à‐vis Turkey. As a result of this defence policy change, military spending increased tenfold in real terms between 1986–99 while as a share of GDP it rose from a yearly average of around 1.8% during 1975–86 to about 3.9% for 1987–99. In order to highlight further the importance of this change, it is interesting to observe that defence expenditure in 1974, the year of the invasion, increased in real terms by about 45.9% to 2.2% of GDP (from 1.2% in 1973) and reached 2.9% of GDP in 1975. As already mentioned, in 1974 and 1975, GDP fell in real terms by 16.9% and 19%, respectively.

Cypriot military spending and GDP growth

Unit Roots and Stationarity Tests

	ADF test		KPSS test		Zivot–Andrews test
Variable	tµ	tτ	ηµ	ητ	t inf	Model type	Break date
gt	−1.529	−5.538**	0.791	0.152	−9.656**	C	1975
	[0]	[0]			[0]
mexp t	−1.899	−2.969	0.329	0.093	−5.052**	A	1986
	[0]	[0]			[1]
Δgt	−9.084**	−9.913**	0.093	0.097
	[0]	[0]
Δmexp t	−6.711**	−6.617**	0.044	0.037
	[0]	[0]
Notes: (i)* and ** indicate significance at the 95% and 99% confidence levels, respectively. (ii) t µ and t τ are the standard augmented Dickey–Fuller test statistics when the relevant auxiliary regression contains a constant, and a constant and a trend, respectively. The number of lagged differenced terms required for serial correlation correction in the ADF auxiliary regressions is selected on the basis of a general to specific testing strategy which is terminated when a sequence of t‐ratio elimination tests on the lagged differenced terms leads to a rejection at the 10% significance level and the residuals of the resultant specification satisfy standard misspecification testing (Ng and Perron, 1995). The selected lag order appears within square brackets underneath the statistics. The response surface regressions of MacKinnon (1991, 1995) are used for determining the significance of the ADF test statistics. (iii) η µ and η τ are the KPSS test statistics for level and trend stationarity respectively (Kwiatkowski et al. 1992). For the computation of theses statistics a Newey and West (1987) robust kernel estimate of the ‘long‐run’ variance is used. The kernel estimator is constructed using a quadratic spectral kernel with VAR(1) pre‐whitenning and automatic data‐dependent bandwidth selection (see, Newey and West, 1994 for details). The asymptotic and finite sample critical values for these tests are taken from Table 1 in Sephton (1995). (iv) t inf stands for the Zivot and Andrews (1992) test statistic for evaluating the presence of a unit root under the null against a shifting mean and/or segmented trend alternative. The selection of the appropriate Zivot‐Andrews auxiliary regression to apply is based on the estimation of the most general specification possible, that is Model C in the Zivot–Andrews terminology, and subsequent evaluation of the significance of the shifting mean and changing trend dummy variables entering the auxiliary regression. Selection of the lag differenced terms in the auxiliary regression follows the Ng and Perron (1995) criterion as in the case of the standard ADF tests above. Asymptotic critical values for the test are taken from Tables 2 and 4 in Zivot and Andrews (1992).

Having established that g_t and mexp_t are integrated of the same order, we now proceed to test whether the series in question move together in the long‐run; that is, whether they share a common trend and hence there exists a long‐run cointegrating relationship. For this to be established, it is necessary that the residuals of the cointegrating regression are stationary (i.e. I(0)). In Tables 2– 5, cointegration tests based on the multivariate cointegration technique developed by Johansen (1988, 1991) and extended by Johansen and Juselius (1990, 1992), are reported starting with the determination of the cointegration rank index, r. The lag length selection criterion, which was chosen to be applied to the two bivariate VAR systems is the Sims (1980) sequential small sample modified likelihood ratio test. On the basis of the results reported in Table 2, the VAR (1) case is chosen as the appropriate lag length for our two bivariate relations. More analytically, in all the examined cases, the LR test as well as the other test criteria qualify the VAR (1) bivariate system. It can also be noted that, according to the results in Table 3, the conditional VAR model is well‐specified, except for the presence of non‐normality.1 However, the asymptotic properties of Johansen's method only depend on the i.i.d. assumption of the errors, and thus the normality assumption is not such a serious problem.2

VAR Lag Order Selection Criteria

Lag	LogL	LR	FPE	AIC	SC	HQ
0	−154.0730	−	75.93764	10.00456	10.27939	10.09566
1	−141.6727	20.92556*	45.10302*	9.479541*	9.937584*	9.631369*
2	−139.6091	3.224384	51.37663	9.600566	10.24183	9.813125
Notes:* indicates lag order selected by the criteria. LR is the sequential modified LR test statistic (each at 5% level of significance), FPE is the Final Prediction Error, AIC is the Akaike information criterion, SC is the Schwartz information criterion, HQ is the Hannan–Quinn information criterion.

Residual Misspecification Tests of the Model with k=1

Eq.	σϵ	LB(16)	ARCH(1)	Skew.	Ex.Kurt.	NORM(1)	R ^2
Δm exp t	0.854	11.654	0.381	0.443	4.322	6.429*	0.334
Δgt	0.949	12.472	0.079	−1.515	9.794	22.579**	0.717
Notes: LB(v) is the Ljung‐Box test statistic for residual autocorrelation, ARCH(v) is the test for heterosckedastic residuals, and NORM(v) the Jarque–Bera test for normality. All test statistics are distributed as χ^2 with the degrees of freedom given in parenthesis. * (**) denotes significance at the 5% (1%) level.
Multivariate residual diagnostics
LB(8)	LM(1)	LM(4)	χ^2(4)
19.668	10.560	0.431	30.132
(0.88)	(0.03)	(0.98)	(0.00)
Notes: LB is the multivariate version of the Ljung‐Box test for autocorrelation based on the estimated auto‐and‐cross‐correlations of the first [T/4=8] lags. LM(1) and LM(4) are the tests for first and fourth‐order autocorrelation distributed, and as c^2 with 28 degrees of freedom is a normality test which is a multivariate version of the Shenton‐Bowman test. Numbers in parentheses refer to marginal significance levels.

The results of the formal tests for the determination of the cointegration rank in our system are presented in Table 4. The diagnostic tests fail to show significant serial correlation. The results indicate the presence of non‐normal residuals but, as indicated by Gonzalo (1994), the performance of the Johansen method is still robust even when the errors are non‐normal. In addition to the formal tests, Juselius (1995) offers further insight into the I(1) analysis as well as the correct cointegration rank. She argues that the results of the trace and maximum eigenvalue test statistics of the I(1) analysis should be interpreted with some caution, suggesting the use of the roots of the companion matrix as an additional tool for determining the correct cointegration rank. As seen from Table 5, the first root is near the unit circle while the second root is substantially smaller and inside the unit circle. This result further supports the choice of one unit root and one cointegration vector in the time‐series process. Finally, recursive tests for parameter stability yielded findings strongly in favour of the sample independence of the cointegration results reported above.

Johansen–Juselius cointegration tests

			95% critical values
r	Trace	λ max	Trace	λ max
r = 0	43.4992*	38.0916*	17.8600	14.8800
r ≤ 1	5.4076	5.4076	8.0700	8.0700
Notes: * Denotes statistical significance at the 5% critical level. r denotes the number of eigenvectors. Trace and λ max denote, respectively, the trace and maximum eigenvalue likelihood ratio statistics. The 95% critical values are taken from Osterwald‐Lenum (1992, Table 1*). A structure of one lag was chosen according to the lag order selection criteria presented in Table 3. A model with an unrestricted constant in the VAR equation is estimated according to the Johansen (1992) testing methodology. A small sample adjustment has been made to the Trace and λ max test statistics, as suggested by Reimers (1992).

The Roots of the Companion Matrix

Real	Complex	Modulus
0.99	0.00	0.99
0.76	−0.00	0.76
Notes: The table shows the real and complex parts as well as the modulus of the estimated p × k roots of the companion matrix from the VAR system, p is the number of variables and k is the number of lags of the VAR.

The next step in our analysis is to test for the presence of causality. For this, the VEC specification for mexp_t and g_t takes the following form:

and is estimated by OLS applied to each equation. The error correction term is formed from the residuals of the cointegrating relation. The optimum lag length k for specifying the VEC has been determined by the joint use of a 10% level sequential F‐type test on the joint significance of the highest order lagged terms in the VEC coupled with extensive diagnostic testing of the residuals. Given the annual frequency of the data set, the maximum lag length was initially set at 2. However, irrespective of the initial maximum lag length chosen, the model consistently failed the Jarque–Berra normality test pointing to the presence of outliers in the residuals.

This means that there is a need to incorporate two impulse dummies for 1975 and 1986 in order to achieve a congruent specification. The dummy variables can be thought of as accounting for the lagged response of the variables under investigation to sample specific events; namely, the 1974 invasion of Cyprus by Turkey and the 1986 shift in defence policy, with the concomitant impact on military expenditure and arms procurement from abroad since no domestic arms production capabilities exist. The resultant VEC model is then used as a platform for assessing the null hypothesis of no Granger causality. This hypothesis is evaluated by two standard Wald χ ² tests performed on each equation of the system. The empirical findings point to the presence of bi‐directional instantaneous causality between military expenditure and the growth rate in Cyprus for the 1964–99 period. Initially, note that both the Δm exp _{t −1} term in equation (1) as well as the Δg _t−1 term in equation (2) are found to be statistically insignificant by the computed Wald tests. From Table 6 it can be easily verified that the χ ² tests assessing the null hypothesis H ¹ ₀:δ*₁ = 0 and H ² ₀:θ*₁ = 0 have a p‐value of 0.205 and 0.380 respectively, indicating that the change in defence spending does not respond to the lagged change in the spending of its adversary. Nevertheless this implied lack of causal feedback between the variables of interest is reversed when the significance of the error correction terms in each equation, that is H ³ ₀:ρ ₁ = 0 and H ⁴ ₀:ρ ₂ = 0, is examined. The p‐values of the associated Wald tests turn out to be 0.793 and 0.00, respectively.

The OLS Estimated VEC Specification and the Granger tests for Temporal Causality

	Estimated Coefficients				Granger causality χ ^2 Wald tests
Regression	Constant	π*1	δ*1	ρ 1	δ*1 = 0	ρ 1 = 0
Δm exp t	0.047	−0.209	−0.031	−0.009	1.608	0.399
	(0.179)	(0.188)	(0.027)	(0.017)	[0.205]	[0.528]
Diagnostics: R ^2 = 0.512, S.E. = 1.004, AR =: F (1, 27) = 0.298, H: F(1,32) = 0.876, NORM: χ ^2 (2) = 16.291, RESET: F (1,27) = 0.271
Notes: S.E. is the standard error of the equation, AR stands for the F version of the LM test for serial correlation. H tests for unconditional heteroscedasticity due to the squares of the regressors, NORM is the Jarque–Berra test for normality and RESET is Ramsey's test using the square of the fitted values. The figures in parenthesis are standard errors while those in square brackets are the p‐values levels. The model involves two impulse dummy variables assuming the value of 1 for 1975 and 1986 respectively.
	Estimated Coefficients				Granger causality χ ^2 Wald tests
Regression	Constant	θ*1	ζ*1	ρ 2	θ*1 = 0	ρ 2 = 0
Δgt	−0.442	−0.903	−1.089	−0.621	0.769	19.66
	(1.066)	(1.121)	(0.159)	(0.100)	[0.380]	[0.00]
Diagnostics: R ^2 = 0.639, S.E. = 5.9799, AR =: F (1,27) = 0.250, H: F (1,32) = 0.935, NORM: χ ^2 (2) = 141.94, RESET: F (1,27) = 0.094

CONCLUSIONS

The aim of this paper was to investigate the causality issue between growth and military spending in the case of Cyprus, thus adding to the pool of relevant literature some further empirical findings from a small island economy with a short history of independence during which it has known invasion, war and occupation. Yet, its growth performance and overall development level have contributed to its recent accession to the European Union (Georgakopoulos, 1999; Georgakopoulos and Pashardes, 1999).

The results reported in this study, covering the period 1964–99, suggest the presence of instantaneous bi‐directional causality between the variables involved. Although, on the basis of this analysis it is not possible to determine the exact nature of this relationship, it is possible that high growth rates have allowed the diversion of resources to military uses allowing, for example, the implementation of an extensive upgrading and modernization of the Cyprus National Guard, especially after the mid‐1980s shift in defence policy (Kollias, 2001). A shift that resulted in a substantial increase in defence outlays, from 1.8% of GDP during 1964–86 to 3.9% during 1987–99. This increase in defence expenditure involved the procurement of major weapons systems such as missiles, tanks, armoured vehicles and helicopters from countries such as France and Russia. As a result, according to ACDA estimations, arms imports increased from 0.98% of total imports during 1975–86 to 3.35% during 1987–97. At the same time, given the fact that no domestic defence industrial capacity exists, through which military spending can affect the economy and growth via the aggregate demand channel, one could tentatively suggest that the increased deterrence and defence capabilities have increased the feeling of general security. This, in turn, positively influences economic activity and growth. This tentative explanation of the causality direction from military expenditure to growth (in a bidirectional relationship) can be further supported by the FDI flows that could be used to approximate the feeling of security (or risk) for an economy. On the basis of UNCTAD data, FDI inflows as a percentage of GFCF in Cyprus have steadily increased from 9% and 6% in 1985 and 1986, respectively, to 32% in 1997, 17% in 1998, 46% in 1999, 53% in 2000 and 43% in 2001. Similarly, inward FDI stock increased from 21% of GDP in 1980 to 48% in 2002. Although by no means should this inflow of FDI be mostly or indeed exclusively attributed to increased security as a result of higher defence spending, it may nevertheless be cited as an indicator of the country's security and prospects. Indeed, the country risk score for Cyprus for 2001 in the UNCTAD inward FDI potential index is 0.737 (score scale 0–1) compared, for example, to a score of 0.701 for Greece, 0.373 for Turkey and 0.827 for the UK. This signifies an improvement from 1991 when the score was 0.681 for Cyprus while the scores for Greece, Turkey and the UK were 0.563, 0.281 and 0.844, respectively.

Notes

† We gratefully acknowledge the useful comments and constructive suggestions by the Editor of the journal and an anonymous referee on earlier versions of this paper. The usual disclaimer applies.

* Corresponding author: [email protected]

We were unable to accept, at the 1% level of significance, the presence of autocorrelation and heteroscedasticity was not detected at the same level of significance.

Gonzalo (1994) has shown that the performance of the maximum likelihood estimator of the cointegrating vectors is little affected by non‐normal errors (see also, Banerjee et al. 1993). Lee and Tse (1996) have shown similar results when conditional heteroscedasticity is present.