significantly better developed than SALW data, thanks largely to the work of SIPRI, which has been compiling and updating worldwide annual data going back to 1950.22 Although limited to major conventional weapons, SIPRI “provides the most
painstakingly researched database” available (Brzoska & Pearson 1994: 20) and is an established source in the arms trade literature. The dependent variable for major conventional arms transfers (mctransfer) uses the SIPRI trend-indicator value (TIV),23 which gives a single value to measure the quantity and quality of actual deliveries in a given year (SIPRI 2007: 429).
The TIV is based on an assessment of “the technical parameters of weapons”
transferred in an export-import year and is a value representing quantity and quality assigned from “an index that reflects its value as a military resource in relation to other weapons.” (SIPRI 2007: 429). It is neither a financial value nor a record of payments made,24 though it does factor core weapons price into its scale where information is available (429-30).25 Because transfers can include gifts or aid, as well as sales, by way of a multitude of financing methods – including gifts, barter, discounts, credit, and cash – the TIV measure is substantially more useful in comparing transfers from year to year and country to country (Brzoska 2004: 113; SIPRI 2007). Moreover, it
22SIPRI is the only public source for annual conventional arms trade disaggregated export-import data for non-US exporters. Although there is little quantitative work on the conventional arms trade, researchers often use SIPRI data to illustrate trends in the trade. Post-Cold War examples include:
Durch (2000); Golde & Tishler (2004); Harkavy (1994); Khanna (1992); Kinsella (1994); Sánchez-Andrés (2004); Sanjian (1991, 1998); Wulf (1991).
23I thank Mark Bromley at SIPRI for providing me this data from the SIPRI Arms Transfer Database from April to June 2006. As of June 2007, this database is online, available at http://www.sipri.org.
24 SIPRI argues that identifying an accurate and reliable price figure for weapons is impossible for three reasons: “First, in many cases no reliable data on the value of a transfer are available. Second, even if the value of a transfer is known, in almost every case it is the total value of a deal, which may include not only the weapons themselves but also other items related to these weapons…as well as support systems…and items related to the integration of the weapon in the arms forces. Third, even if the value of a transfer is known, important details about the financial arrangements of the transfer (e.g., credit or loan conditions and discounts) are often unavailable” (SIPRI 2007: 429).
25For a complete description of SIPRI’s method of calculating its TIVs, see SIPRI (2007: 429-30).
can better factor in sales and gifts of secondhand equipment, to which it would otherwise be difficult to assign a monetary value (Durch 2000: 8). Finally, it is a uniform measure across countries and over time and is constantly updated as new information is made available. Thus while the TIV figures cannot be combined with data from other sources, they are extremely useful for major conventional arms
transfer research. No other source presents such comprehensive and consistent data for exporters and importers worldwide, and it can only be hoped that a similar project emerges for SALW transfers in the years to come.
SIPRI data solely covers major convention arms, defined as large weapons with a military purpose, including one of nine categories: aircraft, armored vehicles, artillery, sensors, air defense systems, missiles, ships, engines, or an “other” fulfilling certain qualifications (SIPRI 2007: 428-29). SIPRI consults a wide range of sources in collecting its data, all of which are open to the public: newspapers, books and
reference works, official national and international documents, and periodicals and journals (430). Some records, of course, may require an informed estimate on the part of the researcher and tend to err on the side of being conservative (430). Once again, moreover, it is important to acknowledge that in working with public sources, an absence of a recorded transfer in an export-import-year can indicate either “no transfer” or simply missing data – a so-far undetected transfer. However, it remains the most thorough information available for the global major conventional arms trade.
Research Design
Based on the coding of the dependent variable, the statistical analyses are done with either OLS or logit regression techniques using panel-corrected standard errors.26
26Because the data is presented as dyad-years (exporter-importer-year), panel-corrected standard errors must be used in order to avoid an understatement of errors due to the high number of error parameters involved in panel data, including panel heteroscedasticity and temporal dependence (Beck & Katz
Logit is used for regressions with SALW transfers, which, by necessity, utilize a dichotomous dependent variable.27 OLS can be used only with continuous dependent variables and, therefore, major conventional arms transfers taken from SIPRI’s trend-indicator value (TIV). Thus, two separate sets of regressions are performed with each of the dependent variables, allowing for the comparison of trends with SALW and major conventional weapons.
I then specify different models for each of the two independent variables of interest: human rights and internal conflict. Recent work in political methodology cautions against building “garbage can” or “kitchen sink” models that include a wide array of variables that may, or may not, influence the outcome of interest.28
Researchers should instead carefully select variables based on the theoretical
relationship between the independent variable of interest and the dependent variable, and between the independent variables themselves. Christopher Achen (2002) argues that without a deliberately limited set of independent variables (no more than three, if possible), the effects of variables of interest on the dependent variable will be
obscured and distorted by collinearity, non-linearities, and other problems (443; See also Achen 2005). Results from models featuring unnecessary variables are often fragile and highly contingent on precise specifications, and make it more difficult to explore nuances of the relationship between variables of interest and the outcome (Achen 2005; Berk 2004).
1995). Neglecting these considerations and using the standard Park method with such data, as Beck and Katz (1995) find, can lead to incorrect statistical results and invalid research findings. See also Beck (2001).
27Logistic regression, used with dichotomous dependent variables, tests “whether two variables are linearly related” and calculates “the strength of the linear relationship” if so (Menard 1995: 1). In essence, it indicates “the predicted probability that a case falls into the higher of the two categories on the dependent variable, given its value on the independent variable” (6). In contrast, one of the core assumptions of OLS regression is that the dependent variables are “continuous, unbounded, and measured on an interval or ratio scale” (4).
28See, for example: Achen (2002, 2005); Kadera & Mitchell (2005); Ray (2003, 2005). For a more general discussion, see Berk (2004).
If scholars are to avoid “garbage can” regressions by limiting the number of variables, then which to include? James Lee Ray (2003) suggests a number of criteria for specifying models, in which he argues that confounding variables – not intervening or competing variables – should be included. That is, models should be limited to only those control variables which threaten to have a confounding effect on the independent variable of interest (see also Kadera & Mitchell 2005). Given that each variable of interest will suggest a different set of possible confounders, it is therefore necessary to specify the models for each of the key independent variables separately. As an added check on these findings, a fully specified model including the complete set of
independent variables will follow the more targeted analyses.
Unlike Blanton’s (2000, 2005) research on the United States, the analysis here is performed in a single-stage of arms exports only, rather than a two-stage Heckman model. In a Heckman model, the first step “is the initial gate-keeping stage in which policymakers determine which nations are worthy of [in this case] aid, and a second step in which levels of assistance are determined for the selected set of nations”
(Meernik et al. 1998: 73, emphasis added). However, the accurate use of this model depends on the existence of at least one independent variable relevant to selection stage but not to the amount stage (Achen 1986: 99).29
In the case of arms transfers, it is difficult to argue that variables affecting the selection of recipients do not also affect the level of exports they receive. Anne Sartori (2003) summarizes the problem well and states, “[W]hen theory dictates identical explanatory variables in the two questions, researchers are left with an unhappy choice: to dredge up an extra explanatory variable for the selection equation (leading to specification error if the variable does not belong there) or to identify only from distributional assumptions about the residuals” (112). Indeed, Blanton has trouble with
29See Achen (1986) and Sartori (2003) for more thorough explanations and discussions of these issues.
this restriction in her own research: In one instance, she simply uses the full model for both equations (Blanton 2000).30 In the second, she reverses the model specification requirements by including all independent variables in the amount stage and excluding a pair of tangential variables from the selection stage (Blanton 2005).31 The single-stage model avoids the pitfalls of the two-single-stage analysis described above. It also acknowledges the reality that factors central to the selection stage are typically also important for the amount stage, which in any case are not distinctly separate stages of export decision-making in practice. Decision-makers do not first decide whether a recipient state is an acceptable customer and then adjust the amount of the arms deal in question. Rather, they approve or deny the export license application as a whole.
In addition, the analysis must be able to identify and compare possible variation in states’ behavior over time. If practice does in fact follow policy, export trends will not remain static over time. I therefore run separate regressions for arms transfers for four different time periods: (1) all years, 1981-2004; (2) Cold War years, 1981-1990; (3) all post-Cold War years, 1991-2004; (4) early post-Cold War years, 1991-1997; and (5) late post-Cold War years, 1998-2004, in which regional arms trade agreements and talk of international norms began to proliferate. The results are
displayed side by side, so as to illustrate most clearly the influence of independent variables in each of the given time periods.
30Sartori (2003) notes that Heckman models without the “extra” variable yield results “based only upon the assumptions about distributional assumptions about the residuals rather than upon variation in the explanatory variables. No one who has proposed such an estimator recommends using it with identical explanatory variables in the two equations” (112).
31Blanton (2005) removes variables for Saudi Arabia and GDP in the selection stage but includes them in the amount stage, specifying the model the reverse of the requirements. She provides no explanation for her seemingly puzzling decision to reverse the Heckman model. Moreover, the dummy variable for Saudi Arabia, as Blanton’s model itself acknowledges, is of little value for the selection stage and should therefore be excluded from the model entirely (all variables in the amount stage must be included in the selection stage). Finally, the research in this chapter will show that GDP is a highly influential variable that can significantly change the results of an equation depending on whether or not it is included. Excluding GDP leads to a poorly specified model and misleading results.
For a more fine-grained approach, I use moving windows (or moving
regression) analyses to explore behavior over time (Beck 1983; Swanson 1998).32 Re-estimated models of five-year blocks or “windows” are used for each dependent variable. The models start with 1982, the first year for the lagged independent variables, and add an additional year to each regression until a five-year window has been reached. There are 23 consecutive five-year windows in total, ending in 2004.33 This technique enables a more nuanced analysis and visually maps trends in arms transfer practice over time, focused here on the independent variables of interest. The use of windows, moreover, enables these models to recognize and illustrate better any changing behavior, which in reality would probably not make abrupt alterations from year to year. Rather, as the moving windows analysis is equipped to show, practices would likely evolve gradually (Swanson 1998: 456-7), as standards become more widespread and commonly accepted by states and domestic actors within them.
Statistical Findings
The results of the quantitative analysis suggest that arms transfer behavior can neither clearly nor consistently be classified as “responsible” or “ethical.” Rather, the overall trends indicate some possibly improving – but far from faultless – behavior with regard to human rights and internal conflict. This section will highlight the important and noteworthy trends and evaluate them in the context of the hypotheses outlined above. Key tables and figures will be provided here, with additional results reproduced in Appendix C and a summary table included at the end of the section.34
32For recent applications in political science, see also Adolph (2004); Kayser (2007); Kwon and Pontusson (2005).
33To illustrate, the first five-year window is 1982-1986, the next is 1983-1987, followed by 1984-1988 and so on.
34Additional findings on democracy, which has not been formally included in decision-making criteria, are also available in Appendix C.
Human Rights
The model to explore the effects of states’ human rights records on their receipt of small and major conventional arms contains three independent variables:
GDP per capita as a rough measure of development, democracy, and internal conflict.
Each of these variables may not only have an effect on the outcome of interest – the transfer of SALW or MCW – but may also have confounding effects on states’ human rights conditions and thus our ability to estimate accurately their effect on export decision-making. According to existing empirical research, there is a negative link between human rights and internal conflict,35 while stronger democracy36 and more advanced economic development37 have been shown to positively influence human rights. Omitting one of these variables therefore has the potential to skew the results.
For example, because GDP per capita is affiliated with both positive human rights and arms transfers, its exclusion from the model risks results that over-estimate the
relationship between good human rights and arms transfers.
The human rights variable is disaggregated into dummy variables for each level of human rights score,38 which are highlighted in the regression tables. The level of human rights score with no violations (one or “very good”) is removed from the statistical analysis as the reference category for the four remaining dummy variables.
Because the majority of wealthy democracies fit into this category and, as recipients, are not the focus of study here, this choice does not detract from the analysis. Recall that if exporter state behavior is to reflect choices of “responsible” transfer policies, at the very least, the highest levels of human rights violations (“very bad”) should reveal
35See, for example: Krain (1997); Lopez (1986: 90); Poe and Tate (1994); Skocpol (1979).
36See, for example: Blanton (1999); Bueno de Mesquita et al. (2005); Davenport (1995, 1999);
Henderson (1991); Howard and Donnelly (1986); Mitchell and McCormick (1988); Poe and Tate (1994); Rummel (1995); Sloan (1984).
37See, for example: Davenport (1995); Davenport and Armstrong (2004); Henderson (1991); Mitchell and McCormick (1988); Poe and Tate (1994); Wolpin (1986: 111, 133).
38See Appendix B, Section 6 for a detailed explanation of human rights variable coding and use.
negative coefficients in the results in the most recent years. Exports driven more by normative concerns than image should go further and exhibit negative coefficients for all categories of human rights scores other than the “good” categories.
FIGURE 3.1. Percentage of SALW Transfers of Human Rights Score, Democratic Exporters to Non-OECD Recipients
First, a brief look at the descriptive statistics reveals a largely non-linear pattern to states’ exports of SALW and MCW and their relationship to recipients’
human rights. States with very good and very bad scores tend to receive smaller percentages of arms transfers, while those states with middling scores receive higher percentages. This is more pronounced with democratic exporters, despite the
expectation that they will pay better attention to the laws and norms of the international community.
FIGURE 3.2. Percentage of MCW Transfers by Human Rights Score, Democratic Exporters to Non-OECD Recipients
Figures 3.1 and 3.2 illustrate these relationships (see Figures C.1 and C.2 in Appendix C for full sample findings). Each bar indicates the percentage of dyad-years receiving arms from all dyad-years at that level of human rights score. Note that the data is largely non-linear and so should not be treated as such. It is only in one case – the full sample of SALW transfers (Figure C.1) – that a linear relationship appears, 39 in which transfers decrease as human rights scores worsen. In the other cases, transfers peak in the middle of the scale – “bad” human rights consistently appear to receive the highest percentage outside of C.1 – and decrease as scores both improve and worsen.
In the case of MCW transfers from democracies to non-OECD importers, the level of
39Figure C.2 comes close to showing a linear relationship, as well. The percentages are fairly consistent across category, except that it peaks with “bad” human rights, like Figures 3.1 and 3.2.
exports to the worst human rights performers is actually greater than the level to the best human rights performers. Nevertheless, the percentages of the worst performers are typically on par with or less than the others. Thus it does seem that the worst human rights violators do generally receive lesser, or at least not better, treatment.
TABLE 3.3. Influence of Human Rights on SALW Transfers, Full Sample
Variable 1981-2004
* Significant at the .05 level; ** Significant at the .01 level
Despite these findings in the descriptive statistics, the regression analyses reveal a somewhat different story. For SALW (Table 3.3, 3.4, and C.1), human rights are only significant for select time periods and levels of violations. Surprisingly,
“good” human rights are never significant determinants of export behavior, in either
the full sample (Table 3.3) or that of democratic exporters to their non-OECD trade partners (Table 3.4).40 On the other end of the spectrum, “very bad” human rights for the full sample are only significant in the full range of years and in the 1991-1997 timeframe, not in the 1998-2004 timeframe suggested by new norms and policies.
Even so, it is a positive coefficient, indicating that human rights did play a role in export decision-making, just not a punitive one. Results for the non-OECD sample (Table C.1) show a similar pattern. However, this behavior does not replicate itself for the sample of democratic exporters to non-OECD importers. Here, the results do show a negative, significant coefficient for human rights, but during the 1980s, not in the post-Cold War era. While the Cold War data – especially for SALW transfers – must be viewed with skepticism and the likelihood of exaggerated optimism, this is
nevertheless an interesting twist on expectations.
Human rights emerge as more consistently significant in the average and bad categories, where violations are perhaps less clearly and overtly bad as in the worst category. For all samples, average and bad human rights for the full time period are positive and significant throughout much of the 1990s, with a still positive but smaller – and no longer significant – coefficient in the most recent time period. It is possible, therefore, that exporters have started to clean up their behavior since the late 1990s.
Nevertheless, the fact remains that countries with poor (but not the worst) human rights scores receive SALW with more frequency than those with the best scores.41
40The occasional negative coefficients for “good” human rights simply indicate that it receives less than the reference category, “very good” human rights, which is not surprising and should perhaps be
40The occasional negative coefficients for “good” human rights simply indicate that it receives less than the reference category, “very good” human rights, which is not surprising and should perhaps be