Agricultural activities carried out in the natural environment for thousands of years in harmony with nature have not caused environmental problems and adverse effects on the environment. However, it has become vital to get more products from each unit area to meet the food needs of the rapidly increasing world population. For this reason, artificial elements have increased considerably, which is entering agriculture. This increase has deteriorated the natural environment. As a result, it has made agriculture a sector that causes environmental problems (Altan et al. 2000).

Environmental problems resulting from agricultural activities negatively affect the atmosphere, soil and water resources and biodiversity. Among these environmental problems, greenhouse gases (GHG) have become a global threat because they cause climate change. GHG emissions from agricultural activities constitute approximately 21% of total anthropogenic GHG emissions. GHG emissions from agricultural activities come second after the energy sector. The primary sources of GHG emissions from agriculture are livestock activities, manure management, rice cultivation, crop residues, stubble burning, fertilizers, and pesticides (Ramachandra et al. 2015; Liu ve ark. 2017).

In recent years, environmental pollution and determining the factors causing this pollution have been extensively discussed in the literature. When the literature is examined, it is seen that few studies are looking at the impact of the agriculture sector on GHG emissions compared to the energy, industry, transportation and waste sectors. Most of the studies on the effects of agricultural activities on GHG emissions have been done recently. A large number of studies on agriculture and environmental pollution in Azerbaijan are related to the impact of agricultural activities on soil and water resources. There are very few studies investigating the effect of agricultural activities on the atmosphere or GHG emissions, and these studies provide only theoretical information on the subject. In order to fill this gap in the literature, investigating the relationship between agriculture and GHG emissions in Azerbaijan has been chosen as the subject of the study. Another reason for examining the relationship between agriculture and GHG emissions in this study is that agricultural activities' negative effect on GHG emissions is both a national and international environmental threat.

When the studies on this subject are examined, it is seen that the independent variables that explain the effect of agriculture on GHG emissions are selected in a small number. Therefore, eight variables were chosen as explanatory variables in the study: four agricultural inputs (AI) and four agricultural macro indicators (AMI).

Drabo (2011) investigated the effects of agricultural exports on the environment and human health for 119 countries between 1991 and 2009. The results show that agricultural exports increase agricultural methane (CH₄), nitrous oxide (N₂O) emissions, and water pollution. In addition, it has been determined that environmental pollution caused by agricultural exports negatively affects human health.

Sarkodie and Owusu (2017) examined the relationship between carbon dioxide emissions and livestock and crop production indices in Ghana from 1960-2013. Evidence from the study shows that crop and livestock production index in the long term increases CO₂ emissions. According to the Granger causality test results, a bidirectional causality relationship was found between crop production index and CO₂ emissions. On the other hand, a unidirectional causality relationship from livestock production index to CO₂ emissions was determined.

Appiah et al. (2018) analysed the causal relationship between agricultural production and carbon dioxide emissions in selected emerging economies (Brazil, India, China, South Africa). Fully Modified Ordinary Least Squares (FMOLS) and Dynamic Ordinary Least Squares (DOLS) cointegration techniques were utilized to estimate the long-term relationships between variables in the study. The results of the empirical analysis show that economic growth, crop and livestock production increases carbon dioxide emissions. On the another hand, the findings indicated that the increases in energy consumption and population reduce CO₂ emissions.

Hongdou et al. (2018) investigated the effect of agro-ecosystem on environmental pollution utilizing data from 1960-2014 in China. Various methods were employed for econometric analysis, including unit root test, Johansen cointegration test, Granger causality test, and vector error correction model. The evidence indicates that chemical fertilizer consumption, number of livestock, paddy area, cereal production, stubble burning and agricultural GDP have a statistically significant and positive effect on CO₂ emissions in the long run.

Balogh (2019) analyzed the relationship between agro-ecosystems and carbon footprints for 133 countries employing the panel unit root test and the least-squares estimation method. The results indicate that arable land, number of tractors, fertilizer use, and agricultural exports have a positive and statistically significant effect on the carbon footprint.

Rehman et al. (2019) explored the relationship of GDP, arable land fertilizer consumption, and CO₂ emissions in Pakistan applying the ARDL bound test. Their findings revealed that GDP, arable land, and fertilizer consumption positively and significantly affect carbon dioxide emissions.

Ali et al. (2020) investigated the impact of agricultural production and agricultural energy consumption on carbon dioxide emissions in Ghana by the ARDL bounds test and Granger causality test. According to the results of the ARDL test, crop and livestock production index and agricultural energy consumption have a positive and statistically significant effect on CO₂ emissions in the long run. The Granger causality test results revealed a bidirectional causality between the crop and livestock production index, agricultural energy consumption, and carbon dioxide emissions.

Balogh (2020) studied the effect of AVA, agricultural exports, rice and paddy cultivation area, number of animals, stubble burning on agricultural emissions for 159 countries. The regression analysis results demonstrated that all the variables utilized in the research have a positive and significant effect on agricultural emissions. Stubble burning, livestock, and rice production have a more substantial impact on agricultural GHG emissions than other variables.

Chandio et al. (2020) explored the relationship between forest area, agricultural energy consumption, crop and livestock production, and CO₂ emissions in China from 1990-2016. They used the ARDL bounds test approach, FMOLS, Canonical Cointegration Regression (CCR) long-term estimators, and Granger causality test. According to the cointegration tests, crop and livestock production had a positive effect on CO₂ emissions in both the short and long term. In contrast, agricultural energy consumption and forest areas had a negative impact on CO₂ emissions. Granger causality test results revealed a unidirectional causal relationship running from crop production, agricultural energy consumption, and forest area to CO₂ emission.

3.1. Study Area

The Republic of Azerbaijan, located in the Caspian Sea basin, is the largest country in terms of territory and population compared to other countries in the South Caucasus. Azerbaijan, with a total area of 86.6 million km2, is located in a favorable geographical position and plays the role of a bridge between Europe and Asia. After the collapse of the Soviet Union in 1991, Azerbaijan, which declared its independence, abandoned the closed market model and switched to the open market model. Historically known as an oil country, Azerbaijan also has great potential in agricultural production. After the Contract of the Century was signed in 1994, Azerbaijan managed to export its oil and gas reserves from the Caspian Basin to the European Market and directed the obtained funds to the non-oil sector, mainly to the agricultural sector, for the diversification of the economy. Table 1 shows Azerbaijan's GDP and the share of agricultural production in GDP as the main macroeconomic indicators in different years.

Table 1 . Azerbaijan's GDP and the share of agricultural production in GDP..

Years	Total GDP		Agriculture (% of GDP)
	Value (mln manats)	Gross rate (%)	Value (mln manats)	Gross rate (%)	Agriculture (%)
2000	4718.1		758,9		16,1
2005	12522.5	165,4	1137,9	49,9	9,1
2010	42465.0	239,1	2344,6	106,0	5,5
2015	54380.0	28,1	3359,4	43,3	6,2
2019	81681.0	50,2	4669,6	39,0	5,7

Source: State Statistical Committee of the Republic of Azerbaijan (https://www.stat.gov.az/).

When Table 1 is examined, it is seen that agricultural GDP has increased in value over the years, but decreased in terms of ratio. While the share of the agricultural sector in GDP was 16.1% in 2000, it decreased to 5.7% in 2019. However, this decline in rates should not mean a decrease in output in the agricultural sector.

This decrease is due to reasons such as the rapid development in industry, construction, service, and other sectors, along with the transition to an open market economy. In general, the growth rate of GDP in industry, services, and other sectors in developing countries is higher than the growth rate of the share of the agriculture in GDP.

In Table 2, Azerbaijan is compared with neighboring countries located in the South Caucasus, which have the same history of political and economic development. In the amount of agricultural value added created in the production of agricultural products, Azerbaijan is ahead of the indicators of neighboring countries. Thus, compared to the analyzed the year 2000, in 2019, this indicator increased by 134.5% from 16768.8 million USD to 39328.6 million USD.

Table 2 . Comparative analysis of agricultural CO2 emission and agriculture value added indicators of South Caucasus countries.

Country	2000	2005	2010	2015	2019
Agriculture value added (2015 constant mln.US$)
Azerbaijan	16768,8	23628,0	26685,0	32786,5	39328,6
Georgia	-	10622,1	10121,6	11684,7	12014,5
Armenia	7161,9	11080,2	11267,0	18176,1	14371,2
Agriculture CO2 emissions (mln. kt)
Azerbaijan	43464	41900	30230	35300	37220
Georgia	5698	6051	7124.2	10540	9870
Armenia	3629	4515	4390	5201	6659

Source: WorldBank and FAOSTAT.

According to the amount of CO₂ emission in agriculture, Azerbaijan has the highest carbon dioxide emission in the region. The main reason for this is that Azerbaijan, as the largest country in the South Caucasus, has a production volume greater than the sum of the volume of agricultural production of the other two countries. Despite this, Azerbaijan's CO₂ emissions have been decreasing over the years. Thus, while a similar indicator was 43464 million kilotons in 2000, it decreased by 13.1% to 37220 million kilotons in 2019.

In the end, as a final explanation of both indicators, we can see that the environmental policy of Azerbaijan has a positive result in carbon dioxide emission, and against the background of the increase in the volume of added value created in agriculture, the emission decreases.

3.2. Material

The primary purpose of this study is to examine the impact of selected agricultural inputs and agricultural macro indicators on environmental pollution for the case of Azerbaijan during the period 1992–2018. The annual data utilized in the study were obtained from FAOSTAT and the World Bank database. The variables included in the research were examined in two groups as agricultural inputs and agricultural macro indicators. Each group consists of five variables, one dependent and four independent. Total agricultural greenhouse gas emissions (CO2 equivalent) were chosen as the dependent variable in the research. The independent variables in the AI`s group are agricultural energy consumption, chemical fertilizer consumption, total pesticide use, and the number of livestock. The independent variables in the AMI`s group are agricultural value-added, crop production index, livestock production index and agricultural export index. The explanations of the variables used in the study are presented in Table 3. Figure 1 shows the graph of the time series variables.

Table 3 . Explanatory information about the variables used in econometric analysis.

Variables	Abbreviation	Measurement unit	Source
Agricultural greenhouse gas emissions (CO2 equivalent)	CO2	Gigagrams	FAOSTAT
Agricultural energy consumption	AEC	Terajoule (TJ)	FAOSTAT
Chemical fertilizer consumption	CFC	Ton	FAOSTAT
Pesticide use	PU	Ton	FAOSTAT
Livestock number	LN	Head	FAOSTAT
Agricultural value added	AVA	Constant 2010 US$	World Bank
Crop production index	CPI	Index	World Bank
Livestock production index	LPI	Index	World Bank
Agricultural export index	AEI	Index	FAOSTAT

Figure 1. Graphical representation of series used in empirical analysis.

3.2.1. Econometric model

In order to determine the effect of agricultural activities on environmental pollution, studies in the literature were taken as a foundation. In this direction, a model was created as in equation (1) for AI variables and equation (2) for AMI variables.

(1)$$C{O}_2=f\left(AEC,CFC,PU,LN\right)$$

(2)$$C{O}_2=f\left(AVA,CPI,LPI,AEI\right)$$

The equations (1) and (2) were rewritten and transformed into a linear-logarithmic model.

(3)$$\ln C{O}_{2t}={\beta }_0+{\beta }_1\ln AE{C}_t+\beta 2\ln CFCt+{\beta }_3\ln P{U}_t+{\beta }_4\ln L{N}_t+{\epsilon }_t$$

(4)$$\ln C{O}_{2t}={\beta }_0+{\beta }_1\ln AV{A}_t+\beta 2\ln CPIt+{\beta }_3\ln LP{I}_t+{\beta }_4\ln AE{I}_t+{\epsilon }_t$$

In the model, ln denotes the natural logarithm, β represents the effect ratio of independent variables on the dependent variable, and ε_t is the error term in the model.

3.3. Method

The time-series approach was used to investigate the relationship between agricultural activities and environmental pollution. In the first phase, unit root tests were implemented to examine the stationarity of the series. This study utilizes the Augmented Dickey-Fuller (ADF), Philips-Perron (PP), Kwiatkowski-Phillips-Schmidt-Shin (KPSS) and Zivot and Andrews (ZA) unit root tests. After testing the stationarity, the ARDL bounds test approach was used to investigate the existence and direction of the relationship between the series. In the nex_t step, FMOLS, DOLS, CCR long-term estimators were employed to robust and support the validity of ARDL bound test results. Finally, Granger (1969) causality test was implemented to determine the causality relationship between the variables used in the research. Figure 2 indicates the graphical representation of the analyzes utilized in the study.

Figure 2. Structure of the methodology.

3.3.1. Stationarity in time series: Unit Root Tests

The first concept we face in studies on time series is stationarity. For this reason, firstly, it is required to examine whether the series is stationary or not. The problem of spurious regression may occur in the analyzes made when the series is not stationary (Granger and Newbold 1974). The spurious regression problem causes incorrect results in relationships between variables. Therefore, making the series stationary is very crucial for the research (Gujarati and Porter 2009). The pioneer studies in the literature on unit root tests belong to Fuller (1976) and Dickey and Fuller (1979, 1981) (Elliott et al. 1996). Modern unit root tests are based on the structure of these pioneering studies.

This study implemented three different unit root tests to examine the stationarity of the series, including the Augmented Dickey-Fuller (1981) (ADF), Phillips and Perron (1988)(PP) and Kwiatkowski-Phillips-Schmidt-Shin (KPSS) (1992). Various information criteria are applied to determine lag lengths in unit root tests. In this study, Schwarz Information Criterion (SIC) was preferred.

3.3.2. Cointegration analysis (ARDL bounds test approach)

Different cointegration tests check long-term relationships between variables in time-series studies. Engle-Granger (1987), Johansen (1988), and Johansen and Juselius (1990) tests are the most widely applied cointegration tests in the literature. However, all variables must be stationary at the first level for these tests to be applied. This situation causes some difficulties in practice. This problem has been solved with the ARDL approach developed by Pesaran and Smith (1998), Pesaran and Shin (1999), Pesaran et al.(2001). The ARDL approach has numerous advantages compared to other cointegration tests, such as: applicable regardless of whether the stationarity level of the series is I(0) or I(1); it can be even implemented in a small samples size, unlike other cointegration tests (Tang 2003). The most crucial point to note in this method is that the dependent variable is stationary at the first difference I(1)(Narayan and Narayan 2004). Because of these advantages, the ARDL model was preferred in the study. Equations (5) and (6) represented the ARDL models established for AI and AMI variables, respectively.

(5)$$\Delta C{O}_{2t}={\alpha }_0+{\displaystyle \sum _{i=1}^m{\alpha }_{1i}\Delta C{O}_{2t-i}}+{\displaystyle \sum _{i=0}^m{\alpha }_{2i}\Delta AE{C}_{t-i}+{\displaystyle \sum _{i=0}^m{\alpha }_{3i}\Delta CF{C}_{t-i}}}+{\displaystyle \sum _{i=0}^m{\alpha }_{4i}\Delta P{U}_{t-i}}+{\displaystyle \sum _{i=0}^m{\alpha }_{5i}\Delta L{S}_{t-i}}+{\alpha }_{6}C{O}_{2t-1}+{\alpha }_{7}AE{C}_{t-1}+{\alpha }_{8}CF{C}_{t-1}+{\alpha }_{9}P{U}_{t-1}+{\alpha }_{10}L{S}_{t-1}+{\mu }_t$$

(6)$$\Delta C{O}_{2t}={\alpha }_0+{\displaystyle \sum _{i=1}^m{\alpha }_{1i}\Delta C{O}_{2t-i}}+{\displaystyle \sum _{i=0}^m{\alpha }_{2i}\Delta AV{A}_{t-i}+{\displaystyle \sum _{i=0}^m{\alpha }_{3i}\Delta CP{I}_{t-i}}}+{\displaystyle \sum _{i=0}^m{\alpha }_{4i}\Delta LP{I}_{t-i}}+{\displaystyle \sum _{i=0}^m{\alpha }_{5i}\Delta AE{I}_{t-i}}+{\alpha }_{6}C{O}_{2t-1}+{\alpha }_{7}AV{A}_{t-1}+{\alpha }_{8}CP{I}_{t-1}+{\alpha }_{9}LP{I}_{t-1}+{\alpha }_{10}AE{I}_{t-1}+{\mu }_t$$

where ∆ is the first difference operator of the variables; μ_t is the error term; and m is the optimal lag length; α₁, α₂, α₃, α₄ and α₅ are the short-run coefficients; α₆, α₇, α₈, α₉ and α₁₀ indicate the long-run coefficients.

The ARDL test is based on the F or Wald statistic. The F value obtained from the ARDL model is compared with the critical values calculated by Pesaran et al. (2001) and Narayan (2005), depending on the sample size. The F-statistic is interpreted in three different scenarios. Firstly, if the calculated F-statistic value is less than the lower critical value, then there is no cointegration relationship between the variables. Secondly, a clear interpretation cannot be made when the F-statistic value is between the lower and upper critical values. Thirdly, there is a cointegration relationship between the variables when the F-statistic value is higher than the upper critical value. After confirming the cointegration relationship between the series, estimate long-run coefficients. Equation (7) and (8) represents the long-run ARDL model established for AI and AMI variables, respectively.

(7)$$C{O}_{2t}={\alpha }_0+{\displaystyle \sum _{i=1}^k{\alpha }_{1i}\Delta C{O}_{2t-i}}+{\displaystyle \sum _{i=0}^k{\alpha }_{2i}\Delta AE{C}_{t-i}+{\displaystyle \sum _{i=0}^k{\alpha }_{3i}\Delta CF{C}_{t-i}}}+{\displaystyle \sum _{i=0}^k{\alpha }_{4i}\Delta P{U}_{t-i}}+{\displaystyle \sum _{i=0}^k{\alpha }_{5i}\Delta L{S}_{t-i}}+{\mu }_t$$

(8)$$C{O}_{2t}={\alpha }_0+{\displaystyle \sum _{i=1}^k{\alpha }_{1i}\Delta C{O}_{2t-i}}+{\displaystyle \sum _{i=0}^k{\alpha }_{2i}\Delta AV{A}_{t-i}+{\displaystyle \sum _{i=0}^k{\alpha }_{3i}\Delta CP{I}_{t-i}}}+{\displaystyle \sum _{i=0}^k{\alpha }_{4i}\Delta LP{I}_{t-i}}+{\displaystyle \sum _{i=0}^k{\alpha }_{5i}\Delta AE{I}_{t-i}}+{\mu }_t$$

After the long-run coefficients, we require to predict the short-run coefficients of the ARDL model. Eq. (9) and (10) denote the ECM-based short-run ARDL model created for AI and AMI variables, respectively.

(9)$$C{O}_{2t}={\alpha }_0+{\displaystyle \sum _{i=1}^k{\alpha }_{1i}\Delta C{O}_{2t-i}}+{\displaystyle \sum _{i=0}^k{\alpha }_{2i}\Delta AE{C}_{t-i}+{\displaystyle \sum _{i=0}^k{\alpha }_{3i}\Delta CF{C}_{t-i}}}+{\displaystyle \sum _{i=0}^k{\alpha }_{4i}\Delta P{U}_{t-i}}+{\displaystyle \sum _{i=0}^k{\alpha }_{5i}\Delta L{S}_{t-i}}+{\alpha }_{6}EC{T}_{t-1}+{\mu }_t$$

(10)$$C{O}_{2t}={\alpha }_0+{\displaystyle \sum _{i=1}^k{\alpha }_{1i}\Delta C{O}_{2t-i}}+{\displaystyle \sum _{i=0}^k{\alpha }_{2i}\Delta AV{A}_{t-i}+{\displaystyle \sum _{i=0}^k{\alpha }_{3i}\Delta CP{I}_{t-i}}}+{\displaystyle \sum _{i=0}^k{\alpha }_{4i}\Delta LP{I}_{t-i}}+{\displaystyle \sum _{i=0}^k{\alpha }_{5i}\Delta AE{I}_{t-i}}+{\alpha }_{6}EC{T}_{t-1}+{\mu }_t$$

ECT_t−1 represents the error correction term, in the short-run ARDL model. The ECT coefficient indicates the long-run balance speed of adjustment later than a short-run shock. This coefficient is required to be negative in sign and statistically significant (Pesaran et al. 2001).

Finally, a series of diagnostic tests are performed to test the stability and fits of the data ARDL models, including cumulative sum (CUSUM) and the cumulative sum of squares (CUSUMSQ) tests, heteroscedasticity, normality, serial correlation etc.

3.3.3. FMOLS, DOLS, and CCR long-term estimators

To strengthen and support the validity of the ARDL bounds test results, FMOLS, DOLS, and CCR estimators were used. FMOLS, DOLS, and CCR estimators are frequently employed in research because of their advantages, such as determining long-term relationships among variables, easy interpretation of coefficients, eliminating the internality problem, and giving reliable results in small samples (Adom 2015).

The FMOLS method, developed by Phillips and Hansen in 1990, is an improvement of the least squares (OLS) method and eliminates the diagnostic problems in standard estimators. On the other hand, with the help of the FMOLS method, accurate and unbiased results are obtained in series with a small number of observations (Phillips and Hansen 1990).

The DOLS estimator was first applied by Stock and Watson in 1993 and was developed by Pedroni (2000, 2001). With the help of the DOLS estimator, the long-term coefficients of the independent variables can be estimated and the deviations caused by the internality problem between the error term and the independent variables can be eliminated (Nazlıoğlu 2010).

The CCR method developed by Park (1992) is closely related to the FMOLS method. However, unlike the FMOLS method, the stationary values of the variables are used, not the level values, in order to eliminate the long-term dependence between stochastic shocks and the cointegration equation in CCR. In the CCR method, as in the FMOLS method, the error terms and covariance matrices are obtained first (Küçükaksoy et al. 2015).

3.3.4. Granger causality analysis

Cointegration tests show a causal connection among the variables. However, it does not specify the direction of causality. For this reason, the Granger (1969) causality test is applied to determine the direction of causality between the variable. To implement the Granger causality test, the variables must be both stationary and cointegrated. If the series is not stationary, the Granger causality test is not allowed to be applied (Gokmenoglu and Taspinar 2018). We conducted a pairwise Granger causality test, and the equations are as follows:

(11)$$X_t={\displaystyle \sum _{i=1}^k{\alpha }_i{x}_{t-i}}+{\displaystyle \sum _{i=1}^k{\beta }_i{y}_{t-i}+e_{1t}}$$

(12)$$Y_t={\displaystyle \sum _{i=1}^k{\alpha }_i{y}_{t-i}}+{\displaystyle \sum _{i=1}^k{\beta }_i{x}_{t-i}+e_{2t}}$$

where x_t and y_t are related variables; k is a delay value, e1t and e2t are white-noise error terms.

In the causality test, the null hypothesis is set as H₀:α₁ = 0, which means x_t does not Granger cause y_t. The alternative hypothesis is set as H₀:α₁ ≠ 0, which means x_t does Granger cause y_t. This approach recommends three separate models of causality. The first model is unidirectional causality which means when x_t does cause y_t, but y_t does not cause x_t. Nex_t, the bi-directional causality model means x_t causes y_t and y_t also, cause x_t. The third and final model of the causality displays when x_t is free of y_t and y is free of x_t to mean no causality (Ullah et al. 2018).

4.1. Unit Root Test Results

Before analyzing whether there is any cointegration relationship among the variables, the stationarity of the series should be investigated. The obstacle of spurious regression arises in studies with non-stationary time series. On the other hand, for the ARDL bound test to be applied, the series must be stationary at the maximum first level I(1). For this reason, firstly, the stationarity test was performed with the help of ADF, PP, and KPSS unit root tests. The results of the stationarity test for the AI and AMI variables are presented in Table 4.

Table 4 . Unit root test results.

Level
Variables	ADF		PP		KPSS
	t-Statistic	Prob.	t-Statistic	Prob.	LM-Stat
CO2	0.254	(0.971)	0.006	(0.951)	0.201
AEC	-1.722	(0.409)	-1.722	(0.409)	0.107
CFC	-1.634	(0.452)	-1.634	(0.452)	0.328
PU	0.146	(0.963)	0.021	(0.952)	0.248
LN	-2.745	(0.081)	-0.741	(0.819)	0.171
AVA	0.474	(0.982)	-0.074	(0.943)	0.238
CPI	-0.121	(0.937)	-0.319	(0.909)	0.196
LPI	0.642	(0.988)	0.317	(0.975)	0.213
AEI	-1.265	(0.630)	-1.041	(0.723)	0.126
First Difference
ΔCO2	-4.166*	(0.004)	-4.504**	(0.002)	0.731*
ΔAEC	-5.709**	(0.000)	-5.644**	(0.000)	0.278*
ΔCFC	-6.658**	(0.000)	-6.804**	(0.000)	0.501**
ΔPU	-3.848**	(0.008)	-3.845**	(0.008)	0.647*
ΔLN	-4.537**	(0.002)	-2.744*	(0.031)	0.701*
ΔAVA	-3.622**	(0.003)	-3.741**	(0.010)	0.702*
ΔCPI	-4.483**	(0.002)	-4.582**	(0.001)	0.680*
ΔLPI	-4.863**	(0.001)	-5.458**	(0.000)	0.755**
ΔAEI	-6.780**	(0.000)	-6.881**	(0.000)	0.632*

Note: Proper lag length in ADF, PP, and KPSS tests was determined according to Schwarz information criterion (SIC). *, ** symbols represent 5% and 1% statistical significance levels, respectively..

When the results of Table 4 are examined, it is seen that all variables contain unit roots at their levels according to the results of all three unit root tests. But when the first difference is taken, all the series become stationary. For the ARDL test to be performed, the dependent variable (CO₂) must be the first-order stationary, not at its level. Unit root test results indicate that this condition is provided.

4.2. ARDL Bound Test Result

The prerequisite for the ARDL model is to determine the appropriate lag length. The lag length providing the smallest critical value is defined as the model's optimal lag length. If an autocorrelation problem exists in the created model, then the next lag length provides the smallest critical value is chosen. The AIC information criterion was based on determining the appropriate lag length for the variables. The ARDL models that minimize the AIC information criterion for AI and AMI are reported in Figure 3. According to AIC, it was determined as the appropriate ARDL models for AI variables (3, 1, 3, 0, 1) and AMI variables (2, 1, 0, 0, 1).

Figure 3. The 20 suitable models according to the Akaike Information Criterion for AI (a) and AMI (b) variables.

4.3. Diagnostic Tests for ARDL Models

After selecting the appropriate ARDL model, we applied various diagnostic tests to determine the models’ goodness of fit, including the Breusch-Godfrey LM test for autocorrelation problem, ARCH test for heteroskedasticity problem, Jarque-Bera test for normal distribution and Ramsey RESET test for model establishment error. Table 5 presented the diagnostic test results for both AI and AMI variables.

Table 5 . Diagnostic test results.

Agricultural input variables, model (3, 1, 3, 0, 1)
Testler	Coefficient	Probability
Breusch-Godfrey LM testi	1.2749	0.3469
ARCH testi	0.3501	0.5604
Jarque-Bera testi	2.3836	0.3037
Ramsey RESET testi	1.6181	0.1296
Agricultural macro indicator variables, model (2, 1, 0, 0, 1)
Breusch-Godfrey LM testi	1.0056	0.3908
ARCH testi	1.1511	0.2949
Jarque-Bera testi	3.1884	0.2031
Ramsey RESET testi	1.0331	0.3180

When the results of Table 5 are examined, it is seen that the probability values of all diagnostic tests are higher than the 5% significance level. These diagnostic test results indicate no serial correlation problem, heteroskedasticity problem, normal distribution problem, and model building error in both models.

Another diagnostic test for the ARDL model is the CUSUM and CUSUMSQ tests. The purpose of CUSUM and CUSUMSQ tests is to test the stability of the model (Figure 4). Looking at the results of Fig. 4, it is seen that the test statistics are within the critical bounds at the 5%significance level for both models. This result means that the H0 hypothesis is accepted, and therefore both predicted models are stable.

Figure 4. CUSUM and CUSUMSQ tests for agricultural input (a) and agricultural macro indicator variables (b).

After determining the appropriate lag length for ARDL models and implementing diagnostic tests, it was started to investigate the long-run cointegration relationship among the series. To decide whether there is a long-run cointegration relationship among the variables, it is necessary to determine the F-statistic using the bound test. Bounds test results for AI and AMI variables are presented in Table 6.

Table 6 . Bounds test results for cointegration.

Agricultural input variables
Model	Optimal lag lenght		F-statistic
CO2 = f(AEC, CFC, PU, LN)	(3, 1, 3, 0, 1)		4.30*
	Agricultural macro indicator variables
Model	Optimal lag lenght		F-statistic
CO2 = f(AVA, CPI, LPI, AEI)	(2, 1, 0, 0, 1)		7.03**
Significance	Narayan		Pesaran
	Lower bounds I(0)	Upper bounds I(1)	Lower bounds I(0)	Upper bounds I(1)
%1	4.28	5.84	3.29	4.37
%5	3.06	4.22	2.88	3.87
%10	2.53	3.56	2.56	3.49

Note: * , ** symbols represent 5% and 1% statistical significance levels, respectively..

According to Table 6, the value of the F-statistic, which tests the long-run relationship for AI variables, was found to be 4.30. This value was higher than the critical upper bound value at the 5% significance level provided by Narayan (2005) and Pesaran et al. (2001). This result shows that there is a long-run cointegration relationship among AI variables. The calculated F-statistic for AMI variables is 7.03, higher than the upper bound critical values of Narayan and Pesaran at the 1% significance level. According to this result, the null hypothesis is rejected. The alternative hypothesis is accepted, which means a long-run cointegrated relationship between the variables.

The next step in the ARDL approach is to predict the long and short-run coefficients. The long and short-run coefficients of AI and AMI variables are presented in Tables 7 and 8, in order of.

Table 7 . Long and short-run coefficients of ARDL(3, 1, 3, 0, 1) model for AI variables.

Long-run coefficients
Variables	Coefficients	Std. Error	T-statistic	Probability
LnAEC	-0.0830*	0.0307	-2.7048	0.0205
LnCFC	0.0699**	0.0169	4.1270	0.0017
LnPU	0.0862**	0.0275	3.1332	0.0095
LnLN	0.7827**	0.0655	11.9581	0.0000
C	-4.3526**	1.0605	-4.1043	0.0017
Short-run coefficients
Variables	Coefficients	Std. Error	T-statistic	Probability
D(LnCO2(-1))	-0.2208	0.1643	-1.3436	0.2061
D(LnCO2(-2))	-0.3288	0.1687	-1.9489	0.0773
D(LnAEC)	-0.0296*	0.0100	-2.9775	0.0126
D(LnCFC)	0.0230**	0.0054	4.2927	0.0013
D(LnCFC(-1))	-0.0136	0.0077	-1.7657	0.1051
D(LnCFC(-2))	-0.0072	0.0060	-1.1853	0.2609
D(LnPU)	0.0550*	0.0214	2.5677	0.0262
D(LnLN)	1.2520**	0.3495	3.5818	0.0043
ECT(-1)	-0.6383*	0.2440	-2.6159	0.0240

Note: *, ** symbols represent 5% and 1% statistical significance levels, respectively..

Table 8 . Long and short-run coefficients of ARDL(2, 1, 0, 0, 1) model for AMI variables.

Long-run coefficients
Variables	Coefficients	Std. Error	T-statistic	Probability
LnAVA	-0.1058	0.2215	-0.4779	0.6392
LnCPI	0.3681*	0.1743	2.1119	0.0508
LnLPI	0.3122**	0.0879	3.5510	0.0027
LnAEI	0.0138	0.0122	1.1272	0.2763
C	7.9216*	3.7752	2.0983	0.0521
Short-run coefficients
Variables	Coefficients	Std. Error	T-statistic	Probability
D(LnCO2(-1))	0.3646	0.1837	1.9848	0.0646
D(LnAVA)	-0.2887	0.2200	-1.3124	0.2079
D(LnCPI)	0.3345**	0.1149	2.9101	0.0102
D(LnLPI)	0.2838**	0.0754	3.7617	0.0017
D(LnAEI)	-0.0191	0.0111	-1.7260	0.1036
ECT(-1)	-0.9087**	0.2347	-3.8727	0.0013

Note: *, ** symbols represent 5% and 1% statistical significance levels, respectively..

When the long-term results are examined related to AI variables, it is seen that there is a negative and statistically significant relationship between AEC and CO₂ emissions. This result means that a 1% increase in AEC causes a 0.08% decrease in agricultural greenhouse gas emissions, in the long run, other variables being constant. These results are in line with the results of Appiah et al. (2018) and Chandio et al. (2020), but it differs from that of Ben Jebli and Ben Joussef (2017), Liu et al. (2017), Agboola and Bekun (2019), Ali et al. (2020) and Koshta et al.(2020). Studies examining the relationship between energy consumption and CO₂ emissions have mostly found that energy consumption increases CO₂ emissions. The reason for this the energy used is mostly fossil fuel sourced. The negative impact of agricultural energy consumption on CO₂ emissions in Azerbaijan is possibly due to two reasons. The energy used in agriculture in Azerbaijan originates from hydroelectric and natural gas. Hydroelectric energy is a renewable energy type that has minimal effect on CO₂ emissions. Natural gas is a cleaner type of fuel compared to coal and oil. Chemical fertilizer consumption (CFC) has a positive and statistically significant impact on agricultural greenhouse gas emissions in Azerbaijan. An increase in fertilizer consumption by 1% leads to an increase in agricultural emissions by 0.07%. An expected result is a positive relationship between CFC and agricultural emissions. The chemical fertilizers used in agricultural activities increased by 138% in Azerbaijan between 1992 and 2018. This result is compatible with those obtained by Hongdou et al. (2018), Ullah et al. (2018), and Ronaghi et al.(2018). It was found to have a statistically significant and positive effect on agricultural emissions in the long run regarding pesticide use. Accordingly to this result, a 1%increase in pesticide use turns into a 0.09% increase in agricultural emissions in the long run. This result is in line with the results of Ali et al. (2021). The last one of the AI variables, the number of livestock (LN), also has a positive and significant effect on agricultural emissions. A 1% increase in livestock number leads to a 0.78% increase in agricultural CO₂ emissions. This finding is similar to those of Ullah et al. (2018), Hongdou et al. (2018), Balogh (2020), and Ali et al. (2021). The positive effect of the number of livestock on agricultural emissions is higher compared to other AI variables. This is mainly because livestock-related activities such as enteric fermentation, manure left on pasture, and manure management are the primary sources of agricultural emissions.

The short-run coefficients of agricultural input variables are similar to the long-term coefficient results. According to the short-run coefficients of chemical fertilizer consumption, pesticide use and livestock number positively and statistically significant impact on agricultural emissions. In Azerbaijan, a 1% increase in chemical fertilizer consumption, pesticide use, and livestock number will increase agricultural CO₂ emissions by 0.02, 0.05, and 1.25%, respectively. Agricultural energy consumption has a negative and significant impact on agricultural emissions in the short run and the long run. An increase in agricultural energy consumption by 1% leads to a decrease in agricultural emissions by 0.03% in the short run. In Table 7, the coefficient of error correction term (ECT) related to AI variables is found negative in sign and statistically significant at 5% level. The ECM coefficient is -0.64 indicate that deflection from the short-run equilibrium is corrected by 64% over the long run.

Long-run results related to AMI variables show a negative but insignificant relationship between agricultural value-added (AVA) and agricultural emissions. This result is similar to the study results investigating the relationship between agricultural value-added and CO₂ emissions in Azerbaijan by Gurbuz et al. (2020). However, in the study mentioned above, the effect of AVA is both negative and statistically significant on CO₂ emissions. The crop production index (CPI) has a positive and statistically significant effect on agricultural emissions, and this result implies a 1% increase in the CPI increases agricultural emissions by 0.37%. This result is supported by Sarkodie and Owusu (2017), Appiah et al. (2018), Ali et al. (2020), Chandio et al. (2020), and Leitao and Balogh (2020). The incorrect tillage, irrigation methods, overuse of chemical fertilizers, and pesticides increase the negative effects of crop production on the environment. As shown in Table 7, chemical fertilizers and pesticides positively affect agricultural emissions. The livestock production index (LPI) also has a positive and significant effect on agricultural emissions in the long run, like the CPI variable. A 1% increase in LPI increases the agricultural greenhouse gas emissions by 0.31%. The positive relationship between LPI and agricultural emissions is supported by Sarkodie and Owusu (2017), Appiah et al.(2018), Ali et al. (2020), Chandio et al. (2020), Leitao and Balogh (2020), and Ayyıldız and Erdal (2021). The agricultural export index (AEI), the last of the AMI variables, has a positive but insignificant effect on agricultural emissions. This consequence is in line with Drabo (2011), Ronaghi et al. (2018), Balogh (2019), and Balogh (2020). The agricultural export index increased approximately eight times in Azerbaijan between 1992-2018. Despite this increase, the positive but insignificant impact of AEI on agricultural emissions is an actual result for Azerbaijan`s agricultural sector. The short-run coefficients and signs related to AMI variables support the long-run findings. The coefficient of error correction term (ECT) was found to be negative (-0.91) and significant (0.001). This means that 91% of the deviations that may occur in the short term will disappear in a long time.

4.4. FMOLS, DOLS and CCR Test Results

The long-term relationship between GHG emissions and agricultural inputs and agricultural macro indicator variables was examined by FMOLS, DOLS and CCR methods, except the ARDL test. Table 8 represents the FMOLS, DOLS, and CCR tests for AI and AMI variables.

When the results of Table 8 are examined, it is seen that the coefficients, their signs and significance levels found as a result of FMOLS, DOLS and CCR tests for both agricultural input and agricultural macro indicator variables are very similar to the ARDL boundary test results.

4.5. Granger Causality Test Results

Co-integration tests provide knowledge about whether there is a long-term relationship among the variables, but it does not provide information about the direction of this relationship. For this reason, the Granger causality test was applied to determine the direction of the relationship between the variables. In Table 9, the Pairwise Granger causality test results for AI and AMI variables are presented.

Table 9 . FMOLS, DOLS, CCR cointegration test results.

Agricultural Inputs
Variables	FMOLS		DOLS		CCR
	Coefficent	Prob.	Coefficent	Prob.	Coefficent	Prob.
AEC	-0.0236**	0.0137	-0.0146	0.1231	-0.0298**	0.0120
CFC	0.0421**	0.0000	0.0368**	0.0000	0.0461**	0.0000
PU	0.0542**	0.0006	0.0460**	0.0028	0.0584**	0.0003
LN	0.7553**	0.0000	0.7810**	0.0000	0.7366**	0.0000
C	-4.0252**	0.0000	-4.4273**	0.0000	-3.7310**	0.0000
Agricultural Macro Indicators
AVA	-0.0716	0.5687	-0.0163	0.8947	-0.1107	0.3821
CPI	0.2984**	0.0024	0.2252*	0.0209	0.3302**	0.0032
LPI	0.3573**	0.0000	0.3800**	0.0000	0.3568**	0.0000
AEI	0.0096	0.3069	0.0080	0.4077	0.0128	0.2541
C	7.3012**	0.0028	6.3330**	0.0070	7.9990**	0.0014

Note: * and ** symbols represent 5% and 1% statistical significance levels, respectively..

Table 10 . Pairwise Granger causality analysis results.

Agricultural Inputs				Agricultural Macro Indicators
Null hypothesis	F statistic	Prob	Result	Null hypothesis	F statistic	Prob	Result
AEC → CO2	1.32069	0.2892	NO	AVA → CO2	1.31098	0.2917	NO
CO2 ← AEC	3.68954*	0.0433	YES	CO2 ← AVA	0.74861	0.4858	NO
CFC → CO2	3.59245*	0.0465	YES	CPI → CO2	6.49669**	0.0067	YES
CO2 ← CFC	3.88547*	0.0375	YES	CO2 ← CPI	2.08081	0.1510	NO
PU → CO2	0.10806	0.8981	NO	LPI → CO2	4.65021*	0.0220	YES
CO2 ← PU	3.04321	0.0702	NO	CO2 ← LPI	2.82808	0.0829	NO
LN → CO2	3.73235*	0.0419	YES	AEI → CO2	1.61024	0.2247	NO
CO2 ← LN	1.17447	0.3294	NO	CO2 ← AEI	2.26238	0.1301	NO
CFC → AEC	0.92095	0.4144	NO	CPI → AVA	10.3131**	0.0008	YES
AEC ← CFC	0.37751	0.6903	NO	AVA ← CPI	2.60384	0.0988	NO
PU → AEC	1.52831	0.2412	NO	LPI → AVA	4.58445*	0.0230	YES
AEC ← PU	4.38677*	0.0263	YES	AVA ← LPI	0.47371	0.6295	NO
LN → AEC	6.34356**	0.0074	YES	AEI → AVA	1.56978	0.2327	NO
AEC ← LN	0.20684	0.8149	NO	AVA ← AEI	4.78085*	0.0201	YES
PU → CFC	2.97336	0.0740	NO	LPI → CPI	3.76890*	0.0408	YES
CFC ← PU	0.17219	0.8431	NO	CPI ← LPI	0.86237	0.4373	NO
LN → CFC	3.68601*	0.0434	YES	AEI → CPI	4.69705*	0.0213	YES
CFC ← LN	1.39457	0.2710	NO	CPI ← AEI	6.62688**	0.0062	YES
LN → PU	2.5775	0.1009	NO	AEI → LPI	0.70411	0.5064	NO
PU ← LN	0.1568	0.8559	NO	LPI ← AEI	2.39873	0.1165	NO

Note: * and ** symbols represent 5% and 1% statistical significance levels, respectively. The lag length was chosen as 2 according to the AIC information criterion..

When the Granger causality test results for AI variables are examined, firstly, it is seen that there is a bidirectional causality relationship between chemical fertilizer consumption and agricultural emissions. This result implies that CFC and agricultural emissions affect each other in Azerbaijan. There is unidirectional causality from the livestock number to the agricultural greenhouse gas emissions. Thus, both CFC and LN increase affect agricultural CO₂ emissions because chemical fertilizer and livestock activities use intensively in the agriculture sector. Both results are similar to those obtained by Ullah et al. (2018), Hongdou et al.(2018), Ronaghi et al. (2018), Balogh (2020), and Ali et al. (2021). Furthermore, no causal relationship was found between agricultural energy consumption pesticide use and agricultural emissions. However, a unidirectional causality was found running from PU to CFC. This means that pesticide use indirectly increases agricultural emissions, if not directly. In the AMI group, a unidirectional causality was found running from the crop production index to the agricultural CO₂ emissions. In addition, a bidirectional causality relationship was determined between the animal production index and agricultural emissions. Numerous studies, such as Sarkodie and Owusu 2017, Chandio et al. 2020, Leitao and Balogh 2020, have also confirmed the links among crop production index, livestock production index and agricultural emissions. Another result is a unidirectional causality relationship running from LPI and CPI to agricultural value-added. The LPI and CPI increase agricultural value-added, which means an increase in the share of the agricultural sector in GDP. On the other hand, it should be forgotten that CPI and LPI increase agricultural emissions. Therefore, it is essential to develop policies to reduce agricultural emissions without negatively affecting CPI and LPI. Additionally, no causal relationship was found between agricultural value-added and agricultural export index and agricultural CO₂ emissions.

It is crucial to determine the agri-environmental indicators accurately to build environmentally-friendly agricultural policies. Collecting data on agri-environment indicators is an expensive activity furthermore, it requires comprehensive analysis and examination. A database on agri-environmental indicators has not yet been entirely created in Azerbaijan. It is essential to determine the agri-environmental indicators convenient for the country's conditions and regularly renew the data related to these indicators. A certain amount of time is needed for the establishment of this infrastructure. Therefore, activity should be begun to prepare the necessary infrastructure.

Defining the rights and responsibilities of farmers and landowners in agricultural practices is very important for developing environmentally friendly agriculture. After determining these rights and responsibilities, farmers and landowners must be paid for their environmental services or assume responsibility for paying according to the damage they cause to the environment. Another important point in agri-environmental policies is producers. The producers need to have the necessary knowledge about the agriculture-environment to create and sustain environmentally friendly agricultural policies. Studies show that most producers lack knowledge about fertilization and fight against diseases and pests. It is essential to raise consumers' awareness to increase the demand for agricultural products produced with the ecology and environmentally friendly agricultural practices and sustainability of agri-environmental policies. In addition to creating demand for environmentally friendly agricultural products, it is crucial to ensure consumer confidence. Effective environmentally sensitive agricultural policies should be prepared and applied consistently to prevent the growth of environmental problems originating from agriculture. It is impossible to create environmentally friendly agricultural policies without convenient information and infrastructure. Therefore, it is necessary to know the current environmental problems caused by agricultural activities, the scope and causes, and their interrelationships. No system in Azerbaijan observes the environmental problems caused by agricultural activities and provides sufficient information about them. Appropriate methods and techniques (legal regulations, payments, raising awareness of producers and consumers, education, research, development, etc.) should be developed to prevent agro-environmental problems in Azerbaijan. In this context, the environmentally sensitive agricultural policies of developed countries, especially the European Union (EU), should be taken as an example.

In this study, the effect of the agricultural sector on greenhouse gas emissions was investigated. There is a need to calculate the effects of the agricultural sector on CH4 and N2O emissions for Azerbaijan in the future. Examining the impact of agricultural activities on ecological footprint, another crucial environmental pollution indicator, is an essential issue for future studies.