- Econometric Approach

Our econometric analysis is based on the gravity model of trade augmented with a number of variables that capture the politcal situaton of supplier and recipient and others that are expected to infuence trade in arms for the reasons explained below. The gravity model has been widely used to model bilateral trade fows and is suitable to estmate the efect of specifc economic and politcal factors on trade. It was frst used to estmate trade fows by Tinbergen (1962), extended with theoretcal foundatons by Anderson (1979) and later by Anderson & van Wincoop (2003) taking into account relatve trade costs in the form of multlateral resistance to trade⁴⁶.

We use standard gravity variables, namely GDP and GDP per capita, distance between the country-pair and categorical variables that control for geographical and cultural closeness. In additon, we control for the demand and supply of arms due to conficts in the recipient country, arms embargoes against the recipient country and military pacts and strategical agreements between supplier and recipient.

46 For a thorough descripton of the gravity model of trade, see Chapter V in Feenstra (2004).

In order to estmate determinants of arms transfer correctly, we also need informaton on the industrial capacity of the domestc arms industry and domestc demand for arms. As informaton on both for the tme period of this study is difcult to fnd, we add to the model the share of military personnel divided by the total populaton as a measure of the degree of militarizaton in a society and demand for equipment. Militarizaton can afect supply and demand for arms through various channels. A ceteris paribus higher degree of militarizaton and therefore higher domestc demand for arms, is expected to strengthen the domestc arms industry. This is because the domestc producton of arms and equipment is usually preferred over foreign producton for politcal and strategic reasons. This is especially the case if domestc producton can satsfy the needs of the military. Increased demand for domestc producton then makes exports of arms more likely due to the existence of economies of scale. Concerning the probability to import, a higher degree of militarizaton has an ambiguous efect: if higher domestc demand for arms contributes to the formaton of a compettve arms industry that can satsfy domestc demand, it should lead to a decrease in the imports of arms. Otherwise, demand for foreign arms should be positvely related to the size of the military industry.

Politcal factors, the main focus of this investgaton, are modelled using several variables: frst, we include the level of democracy and our measure of politcal orientaton for supplier and recipient.

Second, we introduce the absolute diference in both dimensions between supplier and recipient.

And third, in order to capture the politcal environment in the region of the exporter and the region of the importer, we include the average value for both variables of all countries that are geographically close. The empirical model is specifed as a probit model to estmate the determinants of the probability that countries i and j agree on a transfer of MCWs:

Pr(transfer_ijt=1∣X)=ϕ(α+X 'β+κi+λj+αt+εijt) , (V.1)

where the dependent variable, transferijt, takes the value one if j placed an order of major conventonal arms in i, or in the case of licensed producton, a licence was issued in year t and zero otherwise. The vector of regressors X is assumed to infuence the outcome and consists of the following variables⁴⁷: ln GDPit and ln GDPjt denote the natural logarithm of the gross domestc product for the supplier and the recipient in year t and ln GDPpcit and ln GDPpcjt, the natural logarithm of gross domestc product per capita for both countries. Trade costs proxied by geographical and cultural distance are measured by the natural logarithm of distance between

47 See Table III.III.1 for an extensive descripton of all variables in the model.

88 V - Politcal Determinants of Internatonal Arms Transfers

capitals of i and j (lnDistanceij); a dummy variable that takes the value one if i and j share a border (Contguityij); a common language (Languageij); or common colonial past (Colonyij) and a variable that takes the value one if i or j, and two if both are landlocked (Landlockedij).

The frst politcal dimension in the model is the level of democracy. We account for the level of democracy for the supplier (Polityit) and the recipient (Polityjt), the absolute diference in the level of democracy between both (Polity_difijt) and the average level of democracy of the surrounding countries for the supplier (Polity_regionit) and the recipient (Polity_regionjt). Our measure for the second dimension politcal orientaton is covered in a similar fashion for the supplier (votewithUSAit), the recipient (votewithUSAjt), the absolute diference between both variables (votewithUSA_difijt) and the average value of surrounding countries of the supplier (votewithUSA_regionit) and the recipient (votewithUSA_regionjt).

The degree of militarizaton is included for the supplier (Militarizatonit) and the recipient (Militarizatonjt). The dummy variable Confictjt indicates involvement of the government of the recipient in a military confict with another party and at least 25 batle-related deaths. Pactijt is a variable that takes the value one if countries i and j have any kind of military or strategic agreement in place in year t. This can either be to remain neutral, a promise not to atack each other, to consult each other if a crisis occurs, or to defend each other. Embargojt takes the value one if a mandatory UN embargo is in place against country j in year t. Besides the nonlinear probit model, we estmate a linear probability model (LMP) that has the advantage of giving a rough but easily interpretable impression of the size of the efect for each variable in the model without calculatng marginal efects.

The main concern regarding the estmaton of equaton (V.1) is that estmates are rendered biased by unobservable heterogeneity that is tme invariant and country specifc or tme varying and common to all countries and correlated with the error term. Given the large number of observatons of over 500,000 in our sample, we have chosen to include fxed-efects by “brute-force” which has the disadvantage of having high demands in terms of computatonal power.

According to Baltagi (2013), the bias is then considerably reduced when having a high number of observatons. In order to control for the bilateral tme-invariant heterogeneity, we employ a separate regression based on the approach by Mundlak (1978) and include in equaton (V.1) the tme averages of the tme variant covariates as additonal explanatory variables.