3.1 Model

In the current literature on the emissions from agriculture, there are four common approaches: the index decomposition method, the bottom-up method, the system optimization, and the econometric approaches (see Xu and Lin (Reference Xu and Lin2017) for further details). In this study, we apply the econometric methods to the extended STIRPAT model to investigate the agricultural emissions' determinants. The STIRPAT model is based on the IPAT model by Ehrlich and Holdren (Reference Ehrlich and Holdren1971), which relates the impacts of human aspects and activities including population (P), affluence (A), and technology (T) on the environment (I). The IPAT model (Ehrlich and Holdren, Reference Ehrlich and Holdren1971) is a mathematical notation of the impacts of human activities on the environment, which can be expressed as:

(1)\begin{equation}I = P \times A \times T.\end{equation}

In this model, the environment (I) must be broadly considered with the inclusion of the physical environment of urban ghettos, the human behavioral environment and the epidemiological environment (Ehrlich and Holdren, Reference Ehrlich and Holdren1971). That is, population growth, population size, population density, resource utilization and depletion, and environmental deterioration must be considered jointly. In the model, I represents the environmental impacts of human activities, which is modeled as a multiplication function of three terms: population (P), affluence (A) (i.e., income level), and technology (T) (i.e., the efficiency of resource utilization). The proposed proxies for the variables in this model are as follows. I could be measured using ecological footprint analysis. Additionally, P represents the population of an area, such as the world, and is expressed in human numbers. A represents the average consumption of each person in the population, which is commonly proxied by gross domestic product (GDP) per capita. As an efficiency factor, the environmental effects of T can vary in many ways. Hence, the unit for T is reliant on the situation to which I = PAT is being applied.

The IPAT model is a heuristic model and cannot be used for estimation. As such, the stochastic version of IPAT was developed by Dietz and Rosa (Reference Dietz and Rosa1997), namely, the STIRPAT model, for empirical testing. The STIRPAT model can be summarized as follows:

(2)\begin{equation}{I_{it}} = {\alpha _{it}}P_{it}^{{\beta _1}}A_{it}^{{\beta _2}}T_{it}^{{\beta _3}}{\varepsilon _{it}},\end{equation}

where I, P, A, and T have the same implications as in the IPAT framework for country i at time t. $\alpha$ represents the country-specific effect. ${\beta _1}$, ${\beta _2}$, and ${\beta _3}$ are the elasticities of the environmental impacts with respect to P, A and T, respectively. The logarithmic form of model (2) for testing takes the following form:

(3)\begin{equation}ln{I_{it}} = {\alpha _{it}} + {\beta _1}ln{P_{it}} + {\beta _2}ln{A_{it}} + {\beta _3}ln{T_{it}} + ln{\varepsilon _{it}}\; .\end{equation}

Because the purpose of our study is to examine the determinants of emissions in the agricultural sector (across countries), equation (3) can be reformulated by integrating the appropriate factors. The equation therefore becomes:

(4)\begin{equation}lnA{E_{it}} = {\alpha _{it}} + {\beta _1}lnPO{P_{it}} + {\beta _2}lnLnGD{P_{it}} + {\beta _3}lnE{I_{it}} + {\varepsilon _{it}},\end{equation}

in which AE is GHG emissions from agriculture; POP is population; GDP is GDP, a proxy for affluence (economic development); and EI is energy consumption, a proxy for the technological progress in emissions reduction (Rafiq et al., Reference Rafiq, Salim and Nielsen2016; Lin and Zhu, Reference Lin and Zhu2017). ɛ is the residual term.

Furthermore, the nature of the agricultural sector implies that there are other potential determinants of agricultural emissions such as trade openness (Ertugrul et al., Reference Ertugrul, Cetin, Seker and Dogan2016) and FDI inflows (Ren et al., Reference Ren, Yuan, Ma and Chen2014; Zhu et al., Reference Zhu, Duan, Guo and Yu2016; Rafindadi et al., Reference Rafindadi, Muye and Kaita2018). Specifically, trade, exporting and importing may facilitate extra economic activities that affect agricultural emissions. By contrast, the FDI inflows may have implications for the technological advancement in the production function, which help improve the efficiency in energy consumption, reducing emissions from the agricultural sector. This supports the pollution halo hypothesis. Nevertheless, the FDI inflows may also have unfavorable effects on the emissions because multinational firms could take advantage of the weak environmental regulations in the host countries (especially developing countries). This is in line with the pollution-haven hypothesis.

This study thus contributes to the scarce literature on agricultural emissions by providing a comprehensive analysis that explicitly considers all potential contributors of agricultural emissions in the baseline model. The drivers of agricultural emission are examined in an augmented version of the STIRPAT model, as in equation (4), by adding a set of additional explanatory variables. Specifically, our study is conducted with an annual panel dataset of 89 countries (table A2a, online appendix)Footnote ¹ and three subsamples – 31 HIEs, 26 UMEs, and 32 l LMEsFootnote ² – from 1995 to 2012. The classification of three subsamples into groups based on country and income level is the same as that of the World Bank's country classification.Footnote ³

For the model in equation (4), this study conducts data analysis in per capita terms by dividing both sides of the equation by the total population.Footnote ⁴ The extended STIRPAT model used in this study then takes the following form:

(5)\begin{align}A{E_{it}} &= {\alpha _0} + \alpha_{1t} A{E_{it - 1}} + {\beta _1}Incom{e_{it}} + {\beta _2}Agr{i_{it}} + {\beta _3}Energ{y_{it}}\notag\\ &\quad + {\beta _4}Trad{e_{it}} + {\beta _5}FD{I_{it}} + {\varepsilon _{it}},\; \end{align}

in which i, t denote country i at year t. AE is the total GHG emissions from agriculture, expressed in per capita terms. Income represents affluence, proxied by GDP per capita. Agri is agricultural value added per capita. Energy is proxied by energy use (kg of oil equivalent) per capita. Trade represents trade openness, proxied by total trade value per capita. All these variables are taken as natural logarithms. With this transformation, the estimated coefficients refer to the elasticities of agricultural emissions to the corresponding factors in the model (Sadorsky, Reference Sadorsky2013). FDI is the ratio of net FDI inflow to GDP to proxy for the capital flow. $\alpha$ and $\beta$ are the constant and coefficients. $\varepsilon$ is the residual term. In addition, export and import values per capita are used to compare the effects of trade openness on agricultural emissions through export and import channels. Finally, two other types of agricultural emissions – CH₄ and N₂O emissions – are employed for comparison purposes.

3.2 Data analysis

All the detailed descriptions of the variables are presented in the online appendix (table A3). Data descriptions of the full sample are presented in table A4, while table A5 refers to the data description of each income–country group. Meanwhile, table A6 reports their correlation matrix.

The dynamic of agricultural activities and emissions of the income groups in the past decades has provided some notable results. The ratio of agricultural-value-added-to-GDP is stable in the case of HIEs, while it decreased for both middle-income groups. In the case of low-income economies, the contribution of agriculture to GDP decreased in the period before 2003; afterward, it fluctuated with a slight increase in the following period (figure A1, online appendix). At the global level, the total emissions from agriculture are stable for the period before 2003 and increase in the period afterward (figure A2, online appendix).

Across the income–country groups, the total emissions increased for the case of upper-middle-income countries, and their total emissions exceeded the emissions from the high-income-country group in 2005 (figure A3, online appendix). The emissions from lower-middle-income and low-income economies presented the same pattern. We can therefore divide the countries into three groups: LMEs, UMEs and HIEs.

In 1995, the five largest emitters in agriculture were China, India, the United States, Brazil and Russia (figure A4, online appendix). Most of these countries, except for Russia, remained the world's largest emitters in 2012 (figure A5, online appendix). Specifically, the total emissions in China and India increased substantially, and the emissions in the United States and Brazil did not fluctuate much.

As shown in online appendix table A4, trade, import and export have strong positive correlations, approximately 0.99. As such, including these three variables altogether in the same equation will probably cause the problem of collinearity in the estimation. Therefore, we estimated the models using trade, export and import separately, one by one, for each equation.Footnote ⁵

The study sample of this research has a relatively large number of cross-sections (N = 89 countries) and a reasonably long time dimension (1995–2012, namely, T = 18 years). To examine the existence of cross-sectional dependence in our study sample, Pesaran's CD test (Pesaran, Reference Pesaran2020) is applied to the lagged variables. The results in table A7 (online appendix) show the existence of cross-sectional dependence in the variables. We therefore conducted the Pesaran (Reference Pesaran2007) CIPS (Z(t-bar)) unit root tests for the variables in levels and first differences.

Except for the total emissions and agriculture value-added variables, which are stationary at the 1 per cent level, and trade openness and export, which are stationary at the 10 per cent level, the remaining variables are nonstationary at the levels. The CIPS unit root tests for the variables in the first differences show that all the variables are stationary at the 1 per cent level. In this case, we recruit different panel cointegration tests: the Kao cointegration test (Kao, Reference Kao1999), the Pedroni cointegration test (Pedroni, Reference Pedroni1999), and the Westerlund cointegration test (Westerlund, Reference Westerlund2005). The results (see table A8, online appendix) indicate strong empirical evidence of long-run cointegration relationships between the variables.

3.3 Estimation method

Because the variables are a mixture of I(0) and I(1) stationarities, in addition to the existence of cointegration, the autoregressive distributed lag model (ARDL model) for panel data is probably the most suitable estimator (Odhiambo, Reference Odhiambo2009; Bildirici, Reference Bildirici2014; Abdullahi et al., Reference Abdullahi, Bakar and Bassan2015).

Dynamic panel data estimation is often conducted by using difference or system GMM estimators to manage endogeneity. However, with the existence of cointegration, these estimators are no longer appropriate methods. Furthermore, in our empirical study, we have cointegration among the variables in combination with the mixture of the I(0) and I(1) variables. In this case, the ARDL is regarded as the most appropriate estimator (Odhiambo, Reference Odhiambo2009; Bildirici, Reference Bildirici2014; Abdullahi et al., Reference Abdullahi, Bakar and Bassan2015). Moreover, the ARDL model allows us to identify short-term and long-term effects by including lags of dependent and independent variables, whether the regressors are endogenous or exogenous (Pesaran and Smith, Reference Pesaran and Smith1995; Pesaran and Shin, Reference Pesaran and Shin1998). This model can thus solve the problem of endogeneity in dynamic panel data. With the potential existence of country fixed effects and time fixed effects, the dynamic fixed effects estimator (DFE) is used for the ARDL model (Asteriou and Monastiriotis, Reference Asteriou and Monastiriotis2004). The DFE-ARDL model first detects the short-run and long-run influences of the regressors, and then deals with fixed effects. Moreover, the bias of this estimator is reduced to zero when the time dimension of data is becoming larger, which is suitable for our case with relatively long time dimensions – that is, 1995–2012 (18 years) and 1995–2008 (14 years).

Overall, the DFE-ARDL model can help manage the problems of endogeneity, heteroscedasticity and fixed effects (Pesaran and Shin, Reference Pesaran and Shin1998), and estimate the short-run and long-run influences of the regressors.Footnote ⁶ Nevertheless, with this approach, we assume that there is only one cointegration relationship among the variables of interest.