*4.2. Model Structure and Estimation Strategy*

In line with the aforementioned empirical studies that deal with the treadmill of destruction theory, we construct our baseline model in implicit functional form as follows:

$$\text{CO}\_2 = \text{f} \,\text{(ME, EN)}\tag{1}$$

where CO2 is carbon dioxide emissions, which are the proxy for the environmental degradation. The right-hand-side variable ME represents the defense burden, which is measured as the ratio of military expenditures to GDP, while EN is the proxy for primary energy use. Accordingly, the baseline specification is constructed in the following equation:

$$\text{CO}\_{2\text{it}} = \beta\_0 + \beta\_1 \text{ME}\_{\text{it}} + \beta\_2 \text{EN}\_{\text{it}} + \mathbf{u}\_{\text{it}} \tag{2}$$

where the indices i and t denote cross-sectional units and time periods, respectively. β<sup>0</sup> is the drift parameter, while β<sup>1</sup> and β<sup>2</sup> are the parameters to be estimated. uit denotes the conventional idiosyncratic disturbance term, which follows the i. i. d. process. In line with the theoretical arguments that support the treadmill of destruction theory, we postulate that the defense burden (MEit) and energy use (ENit) are positively associated with carbon dioxide emissions (CO2it). In other words, the hypothesis could be postulated as in the following:


In this respect, we will estimate the relationship given in Equation (2) using linear and nonlinear panel ARDL methods. In this empirical investigation, we specifically focus on the symmetric and asymmetric effects of the defense burden on environmental degradation. By adding the positive and negative partial sums of the defense burden to the linear ARDL model, we can detect the potential effects on carbon emissions of changes in the defense burden. By performing long-term analysis in the ARDL model, the variables can be integrated at different orders. As pioneered by Shin et al. [56], asymmetric relationships among the variables can be examined within the scope of the ARDL model, which was

introduced by Pesaran et al. [57]. Thus, we assume the presence of symmetric effects using the following error correction model:

$$\begin{aligned} \Delta \text{CO}\_{2i} &= a\_i + \gamma\_{1i} \text{CO}\_{2i-1} + \gamma\_{2i} \text{ME}\_{i,t-1} + \gamma\_{3i} \text{EN}\_{i,t-1} + \sum\_{j=1}^{k} \delta\_{1ij} \text{ACO}\_{2it-j} + \sum\_{j=0}^{k} \delta\_{2ij} \text{AME}\_{it-j} \\ &+ \sum\_{j=1}^{k} \delta\_{3ij} \text{AEN}\_{it-j} + \pi\_{1i} \overline{\text{ACO}}\_{2t} + \pi\_{2i} \overline{\text{CO}}\_{2t-1} + \pi\_{3i} \overline{\text{ME}}\_{t-1} + \pi\_{4i} \overline{\text{EN}}\_{t-1} + \pi\_{5i} \overline{\text{AME}}\_{t} \\ &+ \pi\_{6i} \overline{\text{ACN}}\_{t} + \sum\_{\lambda=2}^{p} \pi\_{7i\lambda} \overline{\text{ACO}}\_{2t-\lambda} + \sum\_{\lambda=1}^{p} \pi\_{8i\lambda} \overline{\text{AME}}\_{t-\lambda} + \sum\_{\lambda=1}^{p} \pi\_{9i\lambda} \overline{\text{ACN}}\_{t-\lambda} + u\_{it} \end{aligned} \tag{3}$$

As argued by Ullah et al. [14], the main advantage of this setup emanates from the combination of the short-term and long-term effects into single equation [15]. In accordance with Eberhardt and Presbitero [58], the short-term effects are captured by the term Δ, which denotes the first differences of the relevant variable, whereas the bar notation denotes the cross-sectional means of relevant variables in Equation (3). The long-term dynamics are captured by the normalization of the estimates of *γ*2*<sup>i</sup>* and *γ*3*<sup>i</sup>* on *γ*1*i*. Following Eberhardt and Presbitero [58] and Ullah et al. [15], to determine whether the defense burden has asymmetric effects on CO2 emissions, we decoupled the defense burden into positive and negative partial sums using the following equations:

$$ME\_{it}^{+} = \sum\_{j=1}^{t} \Delta ME\_{ij}^{+} = \sum\_{j=1}^{t} \max(\Delta ME\_{ij\prime}, 0) \tag{4}$$

$$ME\_{it}^{-} = \sum\_{j=1}^{t} \Delta ME\_{ij}^{-} = \sum\_{j=1}^{t} \min(\Delta ME\_{ij\prime}, 0) \tag{5}$$

where *ME*<sup>+</sup> *it* represents the positive partial sums of the defense burden and *ME*<sup>−</sup> *it* represents the negative partial sums of the defense burden. In this respect, by using Equations (4) and (5), an asymmetric error correction representation of Equation (3) is given as follows:

$$\begin{aligned} \Delta \text{CO}\_{2it} &= a\_i + \omega\_{1i} \text{CO}\_{2i-1} + \omega\_{2i} \text{ME}\_{i,t-1}^+ + \omega\_{3i} \text{ME}\_{i,t-1}^- + \omega\_{4i} \text{EN}\_{i,t-1} \\ &+ \sum\_{j=1}^k \gamma\_{1ij} \Delta \text{CO}\_{2it-j} + \sum\_{j=0}^k \gamma\_{2ij} \Delta \text{ME}\_{i,t-j}^+ + \sum\_{j=0}^k \gamma\_{3ij} \Delta \text{ME}\_{i-j}^- + \sum\_{j=1}^k \gamma\_{4ij} \Delta \text{EN}\_{i1-j}^- + \tau\_{1i} \overline{\Delta \text{CO}\_{2i}} + \tau\_2 \overline{\text{CO}\_{2t-1}} \\ &+ \tau\_3 \overline{\text{ME}^+}\_{t-1} + \tau\_4 \overline{\text{ME}^-}\_{t-1} + \tau\_5 \overline{\text{ME}^-}\_{t-1} + \tau\_6 \overline{\text{EN}\_{t-1}} + \tau\_{6i} \overline{\text{AD} \text{ET}^+}\_t \\ &+ \tau\_7 \overline{\text{AD} \text{ET}\_t} + \tau\_8 \overline{\text{AD} \text{ET}\_{t-1}} + \sum\_{\lambda=2}^p \tau\_{6\lambda} \overline{\text{AD} \text{C}\_{2t-\lambda}} + \sum\_{\lambda=1}^p \tau\_{10\lambda} \overline{\text{AD} \text{ET}\_{t-\lambda}^+} + \sum\_{\lambda=1}^p \tau\_{11\lambda} \overline{\text{AD} \text{ET}\_{t-\lambda}^- + \sum\_{\lambda=1}^p \tau\_{12\lambda} \overline{\text{AD} \text{ET}\_{t-\lambda}^-} + \tau\_{4t} \end{aligned} \tag{6}$$

where *ω*2*i*, *ω*3*i*, and *ω*4*<sup>i</sup>* denote the long-term coefficients, the positive and negative partial sums of the defense burden, and energy use, respectively. In a similar vein, *γ*2*ij*, *γ*3*ij*, and *γ*4*ij* denote the short-term coefficients, the positive and negative partial sums of the defense burden, and energy use, respectively. The existence of the long-term relationship is dependent only if *ω*<sup>1</sup> has a negative value. In order to determine whether asymmetric effects of the defense burden on environmental degradation exist, we test the null hypothesis *ω*2*<sup>i</sup>* = *ω*3*i*. Rejection of the null hypothesis indicates that the effects of the defense burden on environmental degradation tend to be asymmetric in the long term. Likewise, rejection of the null hypothesis *γ*2*ij* = *γ*3*ij* indicates the presence of asymmetric effects of the defense burden on environmental degradation in the short term. For heterogeneous dynamic panel data models, Chudick and Pesaran [59] proposed the dynamic common correlated effects estimator (DCCE), through which we estimated the error correction models shown in Equations (3) and (6). The major superiority of the DCCE estimator lies in the fact that it generates efficient estimates not only in the presence of cross-sectional dependency (CD) and endogeneity, but also in the presence of heterogeneity among the slope coefficients. Furthermore, the consistency of the DCCE estimator stems from the inclusion of the lags of the cross-sectional means of each variable [60].

For investigation of the causal interplay between the variables, both symmetrically and asymmetrically, we employed the panel causality test developed by Dumitrescu and Hurlin [61], in which the CD and heterogeneity of the coefficients for each unit are also considered. The following equation represents the general form of the panel causality test.

$$Y\_{it} = \alpha\_i + \sum\_{k=1}^{K} \gamma\_i^{(k)} Y\_{it-k} + \sum\_{k=1}^{K} \beta\_i^{(k)} X\_{it-k} + \varepsilon\_{it} \tag{7}$$

Along with the estimation of causality between the variables both symmetrically and asymmetrically, the next session will be mainly devoted to the estimation of our NARDL model, represented by Equation (6).

#### **5. Estimation Results**

Through the aforementioned baseline specifications, our empirical analysis consists of four steps. We commence the empirical treatment by checking for slope homogeneity and CD; the relevant results are displayed in Table 3, where the upper part shows the results of homogeneity tests and the lower part shows the results of CD tests. As developed by Pesaran and Yamagata [62], the homogeneity tests firmly indicate the presence of heterogeneity by rejecting the null hypothesis of the homogeneity of the slope coefficients at a 1% significance level. In order to clarify whether cross-correlations among the variables exist, we performed CD and CDLM tests, as introduced by Pesaran [63]. The results of CD tests clearly revealed the existence of cross-correlations among the variables by rejecting the null hypothesis of CD independency at a 1% significance level.

**Table 3.** Slope homogeneity and CD tests.


Notes: significance codes: \*\*\* *p* < 0.01. Source: authors' estimations based on World Bank and British Petroleum (BP) data.

Before proceeding to the symmetric and asymmetric panel ARDL analysis, we performed unit root tests to account for heterogeneity and CD, as pioneered by Pesaran [64]. Table 4 reports the results of the CADF and CIPS tests. The results firmly attest that the series are integrated at different orders. The series of CO2 and ME become stationary by first differencing, whereas the series of EN is stationary at level, i.e., I (0). Thus, as argued by Pesaran et al. [57], it is feasible to utilize ARDL methodology in linear and nonlinear structures.

Table 5 provides the results of panel ARDL and panel NARDL estimates. The coefficient of the error correction term (EC) is negative in each estimate, indicating that there is a cointegration relationship between carbon dioxide emissions and other variables. According to the short-term symmetric panel ARDL estimation results, ΔEN and ΔME are positively associated with CO2. However, in the long term, there is no significant interplay between the defense burden and CO2 emissions, whereas primary energy use has a positive effect on CO2 emissions. A 1% increase in EN exacerbates CO2 emissions by 0.361. The right panel of Table 4 exhibits the estimation results of the panel NARDL model, in which we decoupled ME into positive and negative components to capture the effects of changes on carbon dioxide emissions. We also performed Wald tests so that the coefficients of ME+ and ME<sup>−</sup> were identical to each other to determine the presence of an asymmetric relationship between the defense burden and CO2 emissions. The long-term Wald test statistic (WLR) confirms the asymmetry in the defense burden by rejecting the

null hypothesis, in which the coefficients of ME+ and ME<sup>−</sup> are identical. In this respect, the short-term estimation results demonstrate that positive changes in the defense burden (ΔME+) do not have a significant effect on carbon dioxide emissions. According to WSR, there is no asymmetric relation in the short term. However, negative changes in the defense burden tend to diminish carbon dioxide emissions. Hence, a 1% fall in ΔME− leads to a decrease in CO2 by 0.06%. In line with the symmetric ARDL analysis, the results of the NARDL model verify the distorting effect of rising energy use on environmental quality. A 1% increase in ΔEN tends to increase CO2 by 0.962%. In a similar vein, the long-term asymmetric analysis confirmed the validity the ameliorative effects of a decreasing defense burden on environmental quality. Accordingly, a 1% fall in ME tends to alleviate CO2 by 0.08%. On the other hand, the estimation results points out the positive effect of energy use on CO2 emissions. In this respect, 1% rise in EN causes CO2 to increase by 0.486%.

**Table 4.** Panel unit root tests.


Note: significance codes: \*\*\* *p* < 0.01 and \*\* *p* < 0.05. Critical values at 1%, 5%, and 10% significance levels for both tests are −2.93, −2.76, and −2.66, respectively. Source: authors' estimations based on World Bank and British Petroleum (BP) data.


**Table 5.** Panel ARDL and panel NARDL results.

Note: significance codes: \*\*\* *p* < 0.01 and \* *p* < 0.1. WLR indicates long-term asymmetry test. WSR indicates short-term asymmetry test. Source: authors' estimations based on World Bank and British Petroleum (BP) data.

We finish the empirical analysis by investigating the causal interplay among the variables. The left panel of Table 6 shows the results of the symmetric causality tests, whereas the right panel shows the results of the asymmetric causality tests. The results of the symmetric causality tests demonstrate that unidirectional causality exists from ME to CO2 and EN to CO2, clearly rejecting the null hypotheses at 10% and 5% significance

levels, respectively. To address the asymmetric effects of the defense burden, we examine the causal relationships between the positive and negative components of ME and each variable. In line with the results of the panel NARDL analysis, unidirectional causality is present from ME− to CO2. To this end, the null hypothesis of ME− does not cause CO2 is rejected at a 10% significance level. Finally, it should also be noted that there is no evidence of causality from CO2 and EN to positive or negative changes in ME.

**Symmetric Causality Asymmetric Causality Direction Wbar-Stat. Zbar-Stat. Prob. Direction Wbar-Stat. Zbar-Stat. Prob.** ME → CO2 3.408 6.596 0.063 \*\*\* ME+ → CO2 2.893 5.186 0.148 EN → CO2 3.967 8.126 0.020 \*\* ME<sup>−</sup> → CO2 3.715 7.435 0.067 \*\*\* CO2 → ME 2.411 3.866 0.167 CO2 → ME<sup>+</sup> 1.565 1.548 0.547 CO2 → EN 1.729 1.996 0.512 CO2 → ME<sup>−</sup> 1.429 1.177 0.471 ME → EN 1.204 0.559 0.823 ME+ → EN 1.092 0.252 0.924 EN → ME 2.140 3.122 0.350 ME<sup>−</sup> → EN 1.497 1.361 0.652 EN → ME+ 1.413 1.131 0.724 EN → ME<sup>−</sup> 0.894 −0.287 0.896

**Table 6.** Symmetric and asymmetric panel causality tests.

Note: significance codes: \*\*\* *p* < 0.01 and \*\* *p* < 0.05. Source: authors' calculations based on World Bank and British Petroleum (BP) data.

#### **6. Discussion**

Overall, our findings tend to support the findings of other empirical studies that address this issue for a panel of countries over various time spans [34,35,45–47,50–52]. Furthermore, our findings are in conformity with empirical studies validating the treadmill of destruction theory in the context of individual countries [15,39–41,43]. Thus, in line with the majority of the empirical literature, our findings reveal the validity of the so-called "treadmill of destruction" theory. Aligning with principle of the theory that countries with more labor-intensive and cutting-edge technologies demand more natural resources, our findings show that reductions in military outlays tend to diminish pressure on the environment. Since its establishment, NATO has taken various actions to mitigate pressure on the environment and natural resources due to the potential threat of climate change. The earliest example of this was the establishment of CCMS to initiate and support studies and fellowships to deal with all forms of pollution and the disposal of hazardous waste. These essential actions have been accelerated by the turn of the new millennium, with rapidly growing interest in the climate change and environmental concerns and the initiatives of the UN.

In 2006, CCMS introduced the Science for Peace and Security Plan to execute initiatives dealing with environmental security challenges. Among the most notable examples of this is the Smart Energy Initiative, which promoted energy efficiency and innovative technologies to maintain the operations of the alliance. In addition, the "Green Defense Framework" was ratified by the member states at the Wales Summit in 2014. Since all members of the alliance are involved in both the United Nations Framework Convention on Climate Change and the Paris Agreement, the member states are eager to achieve the target of limiting global warming to 2 or 1.5 degrees Celsius above pre-industrial levels [11,12]. Thus, within NATO, there is a growing tendency to implement innovative and eco-friendly technologies to reduce carbon dioxide and other greenhouse gas emissions through the individual efforts of member states. The majority of members tend to replace obsolete technology with eco-innovative technologies, and tend to reduce the military outlays for the purchase of emissions-producing arms products. Another factor that potentially supports our findings is the restrictions and regulations on carbon emissions within member countries. The countries in the sample are considered to be either high income or upper-middle income. Furthermore, the majority of the sample consists of EU members, for whom regulations on carbon and other greenhouse emissions are relatively strict for producers and suppliers

compared to the rest of the world. Therefore, arms producers are also influenced by regulations that potentially compel the use of eco-innovative technologies.
