ABSTRACT

Our econometric model is based on the neoclassical, long-term growth framework, whereby capital accumulation occurs though savings which are constituted by the fraction of output that is not consumed. Savings in turn depend on GDP as well as on the exogenous population growth (Solow, 1956, 1974). We further follow Dietz et al. (2007) who, in a first attempt to identify the determinants of the level of gross savings (and therefore also genuine savings), provide an overview of the major empirical studies on savings, and conclude that the most significant and robust explanatory variables for genuine savings include GDP per capita and growth. Based on such basic assumptions underlying the neoclassical model and the determinants of savings, we build our initial model as follows:

ANSt ¼ a1ANSt�1 þ a2GDPt�1 þ a3POPt (2)

where ANSt is adjusted net savings at time t that depends on adjusted net savings in the previous period ANSt−1 as well as on GDP and population variables. We then extend this model to include resource dependence, armed conflict, armed violence, and other countryspecific variables. Our analysis is faced with a variety of endogeneity issues. First, descriptive statistics

indicate that the countries under study differ greatly, even within the same income category. Heterogeneity across countries is dealt with by including country-fixed effects. Second, since the data-generating process is dynamic, current observations of the dependent variable depend on past realizations. Therefore, we include the lagged value of the dependent variable in the set of covariates. Third, some explanatory variables such as GDP are clearly correlated with ANS per capita. We follow the standard approach in the literature and tackle the potential endogeneity of the GDP measures by including only the lagged values of per capita GDP and GDP growth, which further accounts for the process of income transformation into capital. Fourth, we address shocks that are time-varying but common across countries such as the increase in commodity prices by including a binary variable for every year. Our basic econometric model then looks as follows:

ANSit ¼ b1ANSi;t�1 þ b2GDPi;t�1 þ b3POPit þ b4COUNit þ b5RDit þ b6V IOLit þ kt þ mi þ it; (3)

where our dependent variable ANSit is per capita genuine savings of country i at time t. The list of control variables includes the lagged dependent variable ANSi,t−1, and the lagged level of per capita GDP, lagged GDP growth, and inflation as denoted by the matrix denoted in current US dollar we also control for inflation in our

DEFENSE SPENDING, NATURAL RESOURCES, AND CONFLICT

analysis. We further include population growth, the percentage of the population living in rural areas, and rural population growth; these variables are all collected in the matrix POPit. The last set of control variables is found in COUNit and contains the surface and the forest area as well as the agricultural value per worker and growth in agricultural value. Our measure of resource dependence is represented by RDit. Armed conflict and homicides are denoted by VIOLit. We decompose the disturbance term as follows: time-fixed effects are captured by λt and country-fixed effects by νi, the idiosyncratic disturbance component is εit. We control for the first two elements of the disturbance term and cluster εit at the country level as we do not expect country-level disturbances to be independent from each other. This specification controls for a wide range of observable country characteristics that are

very likely to be correlated with ANS and deals with time-invariant country-specific endogeneity by including fixed effects. However, we cannot rule out that there are hidden dynamics that jointly affect investment in sustainable development and a country’s exposure to violence, nor can we rule out reverse causality. We thus compare the results from the OLS estimation of equation (3) with the results from a two-step procedure in which we instrument the potentially endogenous variables armed conflict and the homicide rate. In the reduced form regression of the two-step procedure, we regress the endogenous violence variable on the set of covariates of equation (3) and the exogenous instrument:

VIOLit ¼ d1ANSi;t�1 þ d2GDPi;t�1 þ d3POPit þ d3COUNit þ d4RDit þ cINSTit þ ct þ ni þ git; (4)

Finding a time-varying country-specific instrument is a hard task. Prominent and accredited instruments such as settler mortality or ethnolinguistic fractionalization (Easterly and Levine, 1997; Acemoglu et al., 2001) cannot be employed here: the first instrument is timefixed and the second is only considered valid if it relies on past data. Ethnolinguistic composition does not change quickly and thus cannot be exploited here. Therefore, we have to rely on different instruments that identify conflict and violence, and vary over time for any given country. We propose to instrument a country’s propensity for conflict with its total agricultural surface. Agricultural output is a non-negligible component of national income in developing countries. Obviously, this income component has a direct impact on investment in sustainable development and will be immediately controlled for in the structural equation. However, the availability of agricultural land per se does not directly translate into capital accumulation and agricultural yield varies depending on the intensity of agriculture. At the same time, it is is well documented in the literature on land titling and on the relationship between access to land and household income that the availability of agricultural land plays a significant role in developing countries (Feder and Noronha, 1987; Shipton and Goheen, 1992; Jayne et al., 2003). In addition, the extractive resource and food price indices have both followed a similar boom pattern throughout the period under review, making the lack of agricultural land a potential source of conflict as our reduced form results clearly confirm. The instrumentation strategy for the homicide rate differs substantially from the one

employed for armed conflict. It has been argued that a high-population density may be related to an increased risk of homicides (Gillis, 1974; Krahn et al., 1986). However, this is not verified in reality: low homicide rates can be found in societies with very different population density (Neuman and Berger, 1988; Cole and Gramajo, 2009). In turn, a link has been established between rapid population growth and homicides (Messner, 1982; Krahn et al., 1986). Obviously, population growth per se cannot be considered as