The Relationship between Carbon Dioxide Emissions, Economic Growth and Agricultural Production in Pakistan: An Autoregressive Distributed Lag Analysis

: This study aims to explore the casual relationship between agricultural production, economic growth and carbon dioxide emissions in Pakistan. An autoregressive distributed lag (ARDL) model is applied to examine the relationship between agricultural production, economic growth and carbon dioxide emissions using time series data from 1960 to 2014. The Augmented Dickey–Fuller (ADF), Phillips–Perron (PP) and Kwiatkowski–Phillips–Schmidt–Shin (KPSS) tests are used to check the stationarity of variables. The results show both short-run and long-run relationships between agricultural production, gross domestic product (GDP) and carbon dioxide emissions in Pakistan. From the short-run estimates, it is found that a 1% increase in barley and sorghum production will decrease carbon dioxide emissions by 3% and 4%, respectively. The pairwise Granger causality test shows unidirectional causality of cotton, milled rice, and sorghum production with carbon dioxide emissions. Due to the aforementioned cause, it is essential to manage the e ﬀ ects of carbon dioxide emissions on agricultural production. Appropriate steps are needed to develop agricultural adaptation policies, improve irrigation facilities and introduce high-yielding and disease-resistant varieties of crops to ensure food security in the country.


Introduction
The issue of climate change is now a global challenge and has attracted attention of world leaders for proactive and expedited planning for low carbon industrial growth, clean and renewable energy sources, agricultural sustainability and low-level energy-intensive economic growth [1][2][3][4][5][6][7]. To ensure food safety and food security, dedicated actions are needed on climate change and its impacts on food production [3,8,9].
Climate change can affect agriculture productivity through a change in global temperatures, variability in precipitation and other related factors. It is estimated that about 15-30% of the output of agriculture would be affected globally by 2080-2100 [10]. If timely and adequate adaptive measures are not taken, a decline in crop yield may occur in Africa, Latin America and Asia. Further, it would cost about 5-10% of gross domestic product (GDP) for Africa to take adaptation measures to combat climate change. Moreover, the results of the study predicted that about 50% of the decline in agricultural crops

Literature Review
Many previous studies have employed modern econometric techniques to determine the association between environmental greenhouse gasses, energy consumption and socio-economic variables in various nations globally [5,[40][41][42][43][44][45][46][47]. A previous study investigated the relationship between the consumption of electricity, industrialization, GDP and carbon dioxide emissions in Benin using an autoregressive distributed lag (ARDL) model [42]. Evidence from the study revealed a long-run equilibrium association flowing from consumption of electricity industrialization, GDP and carbon dioxide emissions [42]. Another study employed the vector error correction model (VECM) and ordinary least squares (OLS) regression to reveal the impact of population progression, energy intensity and GDP on carbon dioxide emissions in Ghana [48]. The study found evidence of the existence of a long-run equilibrium association flowing from population growth, energy intensity and GDP to carbon dioxide emissions. The study also revealed that there was a bi-directional causality among energy consumption and carbon dioxide emissions [48]. Another study in Ghana investigated the association between population growth, use of energy, GDP and carbon dioxide emissions using both autoregressive distributed lag (ARDL) regression analysis and a vector error correction model (VECM). The study found that there will be fluctuation in carbon dioxide emissions due to the use of energy in the future. Furthermore, evidence from the study showed a unidirectional causality running from carbon dioxide emissions to use the energy and population [49].
Theoretically, an association could be established through microeconomic and macroeconomic dimensions. From the view of the macroeconomic dimension, the two important areas which are stressed include the impact on the output level, such as yields and the ability of the economy to grow [50]. On the microeconomic side, we have factors such as physical productivity of labor, health and conflict. These factors have economy-wide implications [51][52][53]. Moreover, climate change can have such effects as political inconstancy, which may obstruct factor accumulation and growth in productivity [54].
It has been reported that a rise in temperature can have a profound influence on the productivity of the agriculture sector, food security and farmer's income. This effect varies in tropical and temperate areas. In middle and high latitudes, the aptness and output of crops are anticipated to increase and spread northwards, and vice versa is true for several countries in tropical regions [55]. It is found that in high latitudes, production can be increased by nearly 10% due to a 2 • C rise in temperature, while it reduces production by the same percentage in low latitudes. By taking into account the effect of technology, it is projected that an increase in temperature would increase the productivity of yields by between 37% and 101% by the 2050s in the Russian Federation [54].
In comparison to other developing countries, the effects of escalating temperature on agriculture are harsher in sub-Saharan Africa [56]. It has been observed that if some important climatic conditions, such as temperature and rainfall, had persisted at their pre-1960 status, then the gap of agricultural production between different developing countries and sub-Saharan Africa at the end of the 20th century would have remained only 32% of the existing shortfall. A study of the period of 1980-2005 in Nigeria indicated that temperature exerts a negative influence while rainfall has a positive effect on agricultural production [57].
Some illumination of the effects of climate change on African development was provided in the 4th assessment report of the IPCC. For instance, it was estimated that yield could be reduced by 50% by 2020 in some countries, and the revenue generated from crops could fall nearly 90% by 2100. Smallholder farmers would be affected the most. This will also provoke water problems, as almost 25% of the population in Africa has recently encountered high water stress. Because of increasing water stress in Africa, the population at risk is projected to be between 350 and 600 million by 2050 and about 25-40% of mammals may become endangered in national parks in sub-Saharan Africa [11].
Developed countries have the ability to maintain a minimum level of technology for the improvement of living standards and increasing agricultural productivity [58]. These countries are generally also capable of offsetting the negative consequences of climate change. Developed states usually have a low level of susceptibility but a high level of adaptive ability, which itself is a function of technological expertise, dissemination and supply of assets, and human social and political capital [59]. The developed world has good levels of water filtration and sanitation. On the other hand, developing countries have insecure and unreliable water supplies, and often sanitation systems are non-satisfactory. The notion of crop insurance to protect farmers from the negative consequences of climate change, which may destroy their livelihoods, is missing in developing countries.
During the past decade, Pakistan's per capita gross domestic product (GDP) has experienced a diverse trend. During the period from 2005 to 2014, per capita GDP increased from USD 974.5 to USD 1111.2. In 2011, the government placed significant emphasis on upgrading the country's economy, resulting in a consistent increase of per capita GDP during the period 2011 to 2014. During this period, despite several types of socio-economic challenges, such as energy crises, a war against terrorism, and poverty, per capita GDP (Pakistan Economic Survey 2017) increased by USD 64.71, providing evidence that the Government of Pakistan has taken actions to raise economic growth and enriched the living conditions of the hinterlands.

Data Sources and Description
The key purpose of this study is to answer the question: is there any causal effect between carbon dioxide emissions, gross domestic product and agricultural production in Pakistan? The study used time series data from 1960 to 2014. The data for different variables of this study was acquired from Index Mundi and World Development Indicators of the World Bank. Based on the review of literature, the current study uses nine variables: carbon dioxide emissions CO 2 (kt), gross domestic product (GDP) (constant 2010 US$), barley production (1000 Mt), corn production (1000 Mt), cotton production (1000 Mt), milled rice production (1000 Mt), millet production (1000 Mt), sorghum production (1000 Mt) and wheat production (1000 Mt). The trends of the study variables are given in Figure 1.

Econometric Model
Descriptive statistics are estimated to determine the features of the study variables. To find out the integration order of the study variables, in the first step, we have to identify stationarity in the time series data. For this purpose, we employed the Augmented Dickey-Fuller (ADF) [60], Kwiatkowski-Phillips-Schmidt-Shin (KPSS) and Phillips-Perron (PP) unit root tests [61], and the ARDL bounds test was then estimated. Furthermore, the pairwise Granger causality test and variance decomposition analysis were carried out to examine the direction of causality and improve the study variables in the future. Figure 2 presents the schematic diagram of the study.
The econometric specification of the study variables can be written as: The empirical specification of the proposed model is written as: In Equation (2), LnCO 2t is the logarithmic form of carbon dioxide emissions, LnGDP t is the gross domestic product (GDP), LnBARLEY t is the barley production, LnCORN t is the corn production, LnCOTTON t is the cotton production, LnMILLED RICE t is the milled rice production, LnMILLET t is the millet production, LnSORGHUM t is the sorghum production and LnWHEAT t is the wheat production in year t, ε t is the error term, and α 0 , α 1 , α 2 , α 3 , α 4 , α 5 , α 6 , α 7 and α 8 are the elasticities to be estimated in Equation (2).

Descriptive Analysis
The descriptive analysis shows the mean, coefficient of variation, skewness, kurtosis and normality of distribution of the study variables. The results of descriptive statistics of the study variables are estimated in Table 1. Evidence shows that CO2, gross domestic product (GDP), barley, corn, cotton, milled rice, millet and wheat exhibit positive skewness, while sorghum exhibits a negative skewness. The result of the kurtosis test shows that the CO2, gross domestic product (GDP), barley, cotton, milled rice, millet and wheat exhibit a platykurtic distribution, while corn and sorghum exhibit a leptokurtic distribution. The outcome from the Jarque-Bera test shows that we accept the null hypothesis of normal distribution at the 5% level of significance for barley, milled rice, millet, sorghum and wheat crops.

Descriptive Analysis
The descriptive analysis shows the mean, coefficient of variation, skewness, kurtosis and normality of distribution of the study variables. The results of descriptive statistics of the study variables are estimated in Table 1. Evidence shows that CO 2 , gross domestic product (GDP), barley, corn, cotton, milled rice, millet and wheat exhibit positive skewness, while sorghum exhibits a negative skewness. The result of the kurtosis test shows that the CO 2 , gross domestic product (GDP), barley, cotton, milled rice, millet and wheat exhibit a platykurtic distribution, while corn and sorghum exhibit a leptokurtic distribution. The outcome from the Jarque-Bera test shows that we accept the null hypothesis of normal distribution at the 5% level of significance for barley, milled rice, millet, sorghum and wheat crops.

Unit Root Tests
Before estimating the ARDL bounds test co-integration, it is necessary to determine the stationarity of the variables. To meet the stationarity requirement, the study estimates the unit root using the Augmented Dickey-Fuller (ADF) [62], Phillips-Perron (PP) [61] and Kwiatkowski-Phillips-Schmidt-Shin (KPSS) tests in order to have a robust result. The results of the unit root tests are reported in Table 2. The result of the ADF test shows that the null hypothesis of the unit root cannot be rejected at a 5% significance level. The results of the KPSS test show the null hypothesis of stationarity is rejected at a 5% significance level. Evidence from the results of ADF, PP and KPSS unit root tests shows that the series are integrated at I(1).

ARDL Bounds Testing of Co-Integration and Regression Analysis
The current study uses an autoregressive distributed lag (ARDL) bounds testing approach suggested by [63] to determine both short-run and long-run associations of carbon dioxide emissions, gross domestic product and agricultural production. The ARDL bounds testing method is appropriate for those models in which there is a mixture of I(0) and I(1) variables. Another characteristic of this model is that it is appropriate for small sample size, as our sample size is only 54 [63].
It is important to determine how many lags are to be used in an ARDL model. Therefore, to find the optimal number of lags for the model, the unrestricted vector autoregression (VAR) lag selection criteria are tested. Table 3 formulates the lag selection criteria for the model, but the most commonly employed criteria are the Akaike information criterion (AIC) and the Schwarz information criterion (SIC). A previous study used AIC for small sample size [64]. In this study, we employed the Akaike information criterion, which revealed that the most suitable lag value for the model is lag 3. After unit root testing, which showed all variables are integrated at I(1), we carried out the ARDL method of co-integration (bounds testing) to estimate the relationship between the selected variables in this study. The results of the ARDL bounds testing are reported in Table 4. The results indicate that the f-statistic value (4.954551) is greater than the 10% and 5% upper critical values of I(0) bound. The results of the bounds testing validate significant long-run relationships among variables and show  Furthermore, the study uses the Akaike information criterion (AIC) to select the optimal model by employing long-run and short-run associations among variables. Employing the Akaike information criterion shows the top twenty possible ARDL models in Figure 3. Based on the model specification in Equation (2), the short-run and long-run equilibrium relationships of LnCO 2 , LnGDP, Lnbarley, Lncorn, Lncotton, Lnmilledrice, Lnmillet, Lnsorghum and Lnwheat are estimated using the ARDL regression analysis shown in Equation (3).      Table 5 summarizes the results of the short-run equation of the ARDL model. The results show that the speed of adjustment Error Correction Term ECT(−1) value is −0.30225 which shows that there is a long-run and short-run equilibrium relationship running from LnGDP, LnBARLEY, LnCORN, LnCOTTON, LnMILLED RICE, LnMILLET, LnSORGHUM and LnWHEAT to LnCO2. The speed of adjustment is approximately 30.2% in one period of the long-run equilibrium. Table 5 also shows the results of long-run equation results of the ARDL approach. The results of the long-run equilibrium relationship show that a 1% increase in LnBARLEY will decrease LnCO2 by 3%, a 1% increase in LnMILLET will decrease LnCO2 by 0.03%, and a 1% increase in LnSORGHUM  Table 5 summarizes the results of the short-run equation of the ARDL model. The results show that the speed of adjustment Error Correction Term ECT(−1) value is −0.30225 which shows that there is a long-run and short-run equilibrium relationship running from LnGDP, LnBARLEY, LnCORN, LnCOTTON, LnMILLED RICE, LnMILLET, LnSORGHUM and LnWHEAT to LnCO 2 . The speed of adjustment is approximately 30.2% in one period of the long-run equilibrium. Table 5. Short-run and long-run relationship estimates for the selected model ARDL (1,1,3,0,0,0,2,3,3).  2) Table 5 also shows the results of long-run equation results of the ARDL approach. The results of the long-run equilibrium relationship show that a 1% increase in LnBARLEY will decrease LnCO 2 by 3%, a 1% increase in LnMILLET will decrease LnCO 2 by 0.03%, and a 1% increase in LnSORGHUM will decrease LnCO 2 by 3% in short-run estimates. The evidence of the following studies reveals that carbon dioxide emissions increase in the early phases of economic growth and then decline after a threshold point. The findings of these studies (such as [10,[48][49][50][51][52][53][54][55]) show the relationship between carbon dioxide emissions and GDP growth. The findings of previous studies, such as [65] for China, [66] for Tunisia, [67] for Iran, [68] for Pakistan, [69] for Malaysia, [70] for Turkey and [71] for India, indicate that there is a unidirectional causality running from GDP income to carbon dioxide emissions without response, suggesting that emission reduction plans will not restrain trade and industry growth and that the implementation of such plans seems to be a feasible policy strategy in the aforementioned studied countries to accomplish their long-run sustainable growth.

Diagnostic Test
Once the cointegration relationship was confirmed for the different variables, the cumulative sum (CUSUM) and the cumulative sum of the square of the recursive residuals (CUSUM 2 ) were implemented to run the ARDL model in a befitting manner. The CUSUM and CUSUM 2 tests were employed based on the recursive regression residuals as suggested by [72]. Evidence from the cumulative sum (CUSUM) and cumulative sum of squares (CUSUM 2 ) tests show that the plots lie within the 5% significance level. The two straight lines (red color) show the critical bounds at the 5% significant level. The lines (blue color) in the middle represent the measurements for the cumulative sum of the recursive residuals and the cumulative sum of the square of the recursive residuals. The above statements mean that the ARDL model is constant and stable for estimation of the parameters of the ARDL co-integration bounds test, and the long-run and short-run causality relationship. Figure 4 presents the diagnostic and stability tests for the ARDL model and validates the model. of the ARDL co-integration bounds test, and the long-run and short-run causality relationship. Figure  4 presents the diagnostic and stability tests for the ARDL model and validates the model. Several diagnostic tests were undertaken to check for a good fit of the ARDL model. Table 6 shows that the estimation was suitable with regard to serial correlation and heteroskedasticity, and the inverse root of the AR graph shows the stability of the model.  Several diagnostic tests were undertaken to check for a good fit of the ARDL model. Table 6 shows that the estimation was suitable with regard to serial correlation and heteroskedasticity, and the inverse root of the AR graph shows the stability of the model.

Pairwise Granger-Causality Tests
In this study, we applied an ARDL testing model to determine the short-run and long-run relationship between variables. To find out the causality between LnCO2, LnGDP, LnBARLEY, LnCORN, LnCOTTON, LnMILLEDRICE, LnMILLET, LnSORGHUM and LnWHEAT, we used pairwise Granger causality [73] estimations. The results of the pairwise Granger causality test are presented in

Two-Stage Least Square (2SLS) Method for Endogeneity Problem
Endogeneity is a problem when the explanatory variables correlate with the error term. When an endogeneity problem is found in a model or variables, it is resolved by including an instrumental variable. To identify if an endogeneity problem exists, we applied the 2SLS method to the time series data. In the case of endogeneity in the model, there is a need for instrumental variables. We added agriculture value-added (AVA) as an instrumental variable in our model. Table 8 shows the two-stage least square method for the study variables. The model also shows the Durbin-Watson, J-statistic and second-stage results (SSR) for the study variables.

Impulse Response and Variance Decomposition Analysis
Finally, we employed impulse response analysis in which we employ the response of LnCO 2 , LnGDP, LnBARLEY, LnCORN, LnCOTTON, LnMILLED RICE, LnMILLET, LnSORGHUM, and LnWHEAT to explain random innovations among them. The random response is not described by the pairwise Granger causality test. The impulse-response of carbon dioxide emissions to Cholesky One S.D. innovations in other variables are displayed in Figure 5.
for example, GDP, barley production and cotton production, is significant. On the other hand, a one standard deviation shock to GDP causes carbon dioxide emissions to steadily increase within a 10period horizon. Similarly, a one standard deviation shock to barley production causes carbon dioxide emissions to gradually increase within a 10-period horizon, while corn production first increases carbon dioxide emissions over a 2-period horizon, and then starts decreasing over a 10-period horizon. A one standard deviation shock to cotton production causes carbon dioxide emissions to exhibit and up-and-down motion within a 10-period horizon. Figure 6 shows the response of GDP, barley production, corn production, cotton production, milled rice production, millet production, sorghum production and wheat production to carbon dioxide emissions. Table 9 shows the variance decomposition of LCO2, LnGDP, LnBARLEY, LnCORN, LnCOTTON, LnMILLED RICE, LnMILLET, LnSORGHUM and LnWHEAT within a 10-period horizon. The variance decomposition provides evidence of the relative importance of each random innovation in affecting LnCO2, LnGDP, LnBARLEY, LnCORN, LnCOTTON, LnMILLED RICE, LnMILLET, LnSORGHUM and LnWHEAT in the VAR model. This study employed the variance decomposition method, which estimates the percentage of influence of each independent variable on the error variance of the dependent variable [39]. Figure 5 shows that the response of carbon dioxide emissions to corn production, millet production, milled rice production, sorghum production, and wheat production are insignificant within 10-period horizons. On the other hand, the initial response of carbon dioxide emissions to all other variables, for example, GDP, barley production and cotton production, is significant. On the other hand, a one standard deviation shock to GDP causes carbon dioxide emissions to steadily increase within a 10-period horizon. Similarly, a one standard deviation shock to barley production causes carbon dioxide emissions to gradually increase within a 10-period horizon, while corn production first increases carbon dioxide emissions over a 2-period horizon, and then starts decreasing over a 10-period horizon. A one standard deviation shock to cotton production causes carbon dioxide emissions to exhibit and up-and-down motion within a 10-period horizon. Figure 6 shows the response of GDP, barley production, corn production, cotton production, milled rice production, millet production, sorghum production and wheat production to carbon dioxide emissions.   Table 9 shows the variance decomposition of LCO 2 , LnGDP, LnBARLEY, LnCORN, LnCOTTON, LnMILLED RICE, LnMILLET, LnSORGHUM and LnWHEAT within a 10-period horizon. The variance decomposition provides evidence of the relative importance of each random innovation in affecting LnCO 2 , LnGDP, LnBARLEY, LnCORN, LnCOTTON, LnMILLED RICE, LnMILLET, LnSORGHUM and LnWHEAT in the VAR model.

Conclusions and Policy Implications
This study explored the causal relationship between carbon dioxide emissions, economic growth and agricultural production in Pakistan for the time period from 1960 to 2014. By employing the ARDL optimal model, there was evidence of short-run and long-run associations between gross domestic product, barley, corn, cotton, milled rice, millet, sorghum and wheat to carbon dioxide emissions. The evidence from the unit root tests (ADF, PP and KPSS) showed that all study variables are integrated at I(1). The results of the ARDL bounds test showed that there is a co-integration relationship between all the study variables.
The results of the Granger causality test indicated that there is both unidirectional and bidirectional causality between the study variables. The study also applied the two-stage least square method to describe the endogeneity problem in our variables or model. The paper aimed to employ variance decomposition and Cholesky ordering to investigate the future effect of variables on carbon dioxide emissions in the VAR model.
Agriculture plays a very important role and is considered a backbone in a nation's growth. The government of Pakistan is trying to achieve a healthy living style and increase its economic growth. There is a need to improve agricultural productivity through advanced agriculture production techniques. The country is listed among the countries severely affected by climate change [74] despite being a low producer of CO 2 gasses [75] because of its increasing dependence on agriculture for food and fiber needs [76]. The role of extension services is also very important for spreading updated scientific information to farmers.