ABSTRACT
Tables IV and V report the estimation results of the health and education Equations in (1) and (2), respectively. The POLS estimators indicate that military spending is nothing to do with both health and education expenditures. FE and RE specifications lead to a no ‘trade-off’ result, where we control the country-specific heterogeneity to mitigate the omitted variable bias problem. On one hand, it is observed that the coefficient estimates using POLS, FE, and RE methods generally give rise to the same signs as those in terms of various panel GMM estimations. On the other hand, the POLS, FE, and RE may obtain many insignificant parameter estimates (e.g. see Table V). Because POLS, FE, and RE do not treat the right-hand-side variables as endogenous, the estimation results are hence not reliable. In the following discussion, we will concentrate on the results of the consistent parameter estimation using panel GMM method. As pointed out in Cameron and Trivedi (2005), the panel GMM method can be con-
ducted by choosing different lagged variables as instruments. We first consider to include lag one-period (e.g. Milexit�1andWit�1 in Equations (1) and (2)) and two-period (e.g. Milexit�2andWit�2 in equations (1) and (2)) variables as the IV (see GMM1 model in Tables IV and V).10 The lagged variables up to the three periods are also performed in our analysis, indexed by models GMM2-GMM6 in both tables. Overall, instruments based on lagged variables up to two (model GMM1) and three (model GMM2) periods lead to very similar results. When inspecting Table IV, it is clear to see that military spending has a positive impact
on the health expenditure (see the results of models GMM1 and GMM2). The positive crowding-out effect implies that an increase in military spending also promotes the health
DEFENSE SPENDING, NATURAL RESOURCES, AND CONFLICT
spending as % of GDP, suggesting a complementary relationship between the two types of government spending. As for the education equation in Table V, we find that military spending also has a significantly positive impact on education expenditures – once again confirming a positive crowding-out effect of milex vs. welfare spending relationship. This phenomena has been found in previous studies such as Verner (1983), Harris and Pranowo (1988), Frederiksen and Looney (1994), and Kollias and Paleologou (2011). According to Frederiksen and Looney (1994), government is willing to cut from the infrastructure programs instead of curtailing the social welfare expenditures in response to an increase of military spending. In addition, the defense spending can be beneficial to human capital formation since defense personnel and soldiers are appropriately taken care (in terms of health spending) and well trained physically, and receive good skills (in terms of education spending). This is particularly the case since we focus on the well-developed OECD countries. Moreover, even though the military expenditures are mainly paid for new weapons rather than military personnel, the education training programs must be offered for army personnel. Our empirical results are robust across different model specifications in terms of GMM3-GMM6 in Tables IV and V. It is worth noting that our finding that military spending and two social welfare
expenditures (i.e. health and education) are complementary is slightly different from those of Yildirim and Sezgin (2002) and Ali (2011) using the data in Turkey and Egypt, respectively. Their finding that the trade-off is positive between education and defense spending is consistent with our result, while the negative crowding-out effect between health and military expenditures is opposite to our crowding-in effect. The relationship between health and education expenditures in both health and education
equations is distinct. In Table IV, education spending has a positive impact on health spending even though this relationship is not always significant (e.g. models GMM1 and GMM4). However, health expenditures tend to shrink the size of education expenditures as indicated in Table V. Once again, it is not significant in each model specification (e.g. models GMM5 and GMM6). The asymmetric finding between the two social spendings may be due to that the health spending seems to constantly account for a higher rate as % of GDP than education (see Figure 2 and Table II). However, it needs more work to figure out the detailed link between the two variables in the future research. In regard to other control variables, the effect of population on education and health
expenditure is significantly positive in both equations. The evidence may imply that the social welfare expenditures depend on the number of welfare recipients. The positive impact of the tax revenue on education expenditure suggests that the increasing tax revenues can be utilized to finance more social welfare expenditures. The effect of household consumption and government consumption is positive and significant in both equations. However, GDP has a negative effect on the social welfare expenditures, which might indicate that other components (e.g. gross investment and net export) in GDP increase with a faster rate than government spending. To test whether our model specification is right, the Hansen’s over-identifying restric-
tion tests are performed across GMM1-GMM6 in both Tables IV and V. It shows that the p-values are all large enough to support the over-identifying restriction. The weak instrument problem is an issue in an over-identified model. To address this problem, we apply the F-statistic criterion by Staiger and Stock (1997) from the first-stage regression. Tables IV and V show that the panel robust F-statistics (treating military spending as endogenous) are sufficiently large enough to avoid the weak instrument problem in all cases.
