– EMPIRICAL ANALYSIS CAUSAL DYNAMICS BETWEEN DEFORESTATION AND CO2 EMISSIONS IN TANZANIA

(1)

Master Thesis University of Groningen

Faculty of Economics and Business MSc International Economics & Business

08-01-2019

CAUSAL DYNAMICS BETWEEN DEFORESTATION AND CO2

EMISSIONS IN TANZANIA – EMPIRICAL ANALYSIS

ABSTRACT

The dynamic causal relationship between deforestation and CO2 emissions for Tanzania over

the period 1990-2014 is examined. Deforestation as well as economic and population growth are hypothesized to result in increased CO2 emissions. Forests act as carbon sinks and release

carbon dioxide when cleared for agricultural purposes or fuelwood collection. After employing the ARDL bounds test for cointegration, this study finds support for the strong causal relationship between deforestation and CO2 emissions, while economic development has only

a moderate impact. Furthermore, the results suggest that population growth results in agricultural expansion, which simultaneously leads to deforestation.

Keywords: deforestation, carbon emissions, environmental Kuznets curve, Kaya identity, Tanzania, ARDL cointegration

Author: Nikoliya Kovaleva Student number: 3560279

(2)

List of figures and tables

Figure 1 - Environmental Kuznets Curve ... 6

Figure 2 - Conceptual framework... 14

Figure 3 - Decomposition using the Deforestation variable ... 19

Figure 4 - CUSUM squared graph ... 23

Figure 5 - CUSUM squared graph (Deforestation as dependent variable) ... 25

Table 1 - Total CO2 Emissions ... 12

Table 2 - Decomposition of CO2 emissions using the original decomposition equation ... 19

Table 3 - Error Correction representation of ARDL model (CO2 emissions as dependent variable) ... 24

Table 4 - Agricultural productivity ... 25

Table 5 - Error Correction representation of ARDL model (Deforestation as dependent variable) ... 26

Table 6 - Error Correction representation of ARDL model (GDP per capita as dependent variable) ... 27

List of Abbreviations

CO2 - carbon dioxide

FAO - Food and Agriculture Organization of the United Nations GHG - green house gas

IRENA - International Renewable Energy Agency IPCC - Intergovernmental Panel on Climate Change LULUCF - Land Use, Land Use Change and Forestry

(4)

1 Introduction

Tanzania is the most populous country in Eastern Africa (57.31 million people in 2017) with population growth of 2.75% in 2017 compared to the previous year (worldbank.org). Population distribution is extremely uneven with 67% living in the rural areas in 2017 (Cia.gov, 2018). With US$936.33 GDP per capita in 2017, the country is one of the poorest economies (Teske et al., 2017; worldbank.org, 2018). Nonetheless, in the years between 2009 and 2017 GDP growth amounted to 6-7% on average (Cia.gov, 2018).

These figures demonstrate that Tanzania faces a quickly growing population and accelerating economic growth. The country also aims to become a middle-income country in the next decade (IRENA, 2017). However, there are quite some challenges that have to be tackled. Lack of access to electricity in rural areas is one of them. It is widely recognized that energy is an essential part of a sustainable development of a country and it needs to be accessible and affordable. According to IRENA Director-General Adnan Z. Amin (2017), the country ‘has recognized renewables as an important means to meet these challenges and achieve a sustainable energy future’. Tanzania possesses enormous renewable energy resources, that range from hydropower and bioenergy, to geothermal, wind and solar. However, this potential has not been fully exploited and currently only bioenergy is exploited as an energy source (IRENA, 2017). There are several initiatives, which aim to boost the investments in renewable energy such as the establishment of the rural energy agency, subscription to Sustainable Energy for All (SE4All) and encouragement of the private sector (Ahlborg and Hammar, 2014; IRENA, 2017).

(5)

and the Paris Agreement1_{. In Tanzania, however, greenhouse gas (GHG) emissions are mainly}

driven by land use and forestry instead of fossil fuels from manufacturing and transportation since around 30% in value added of GDP are from agriculture, forestry and fishing (AFF) and biomass is the main energy source (worldbank.org, 2018). Therefore, for a developing country such as Tanzania the focus on fossil fuel energy consumption as an explanatory factor for environmental degradation might not be appropriate (at least in the current stage of its development). Instead, deforestation that ultimately results from extensive biomass consumption and land use for agricultural purposes might be the main driving force of CO2

emissions.

There are conceivable climate risks for Tanzania stemming from increased GHG emissions that might affect water and food security and lead to extreme weather events. In order to understand how to best mitigate the effects of GHG emissions, it is essential to understand what drives these emissions. Only then can viable policy solutions be proposed. These might include increased deployment of low-/no-carbon energy, more efficient use of energy or reduced deforestation (IPCC, 2014).

To sum up, Tanzania is a country with a fast population growth and GDP per capita that is rising. Furthermore, a large amount of people is still lacking access to electricity and use biomass for cooking. All this indicates that Tanzania’s share of global emissions will increase in the future. Therefore, it is important to examine the drivers of CO2 emissions in order for

the decision makers to better understand the path of sustainable development. In the following thesis the driving forces of CO2 emissions will be analyzed and compared using two different

approaches: the Kaya identity and the testing of a model based on the Environmental Kuznets Curve literature using the ARDL methodology. This approach sheds an integrated view on the two research streams that have been examined relatively isolated from each other in the past but actually ask a very similar question: ‘does economic growth need to be slowed, if not stopped, in order to avoid increasing harm to the environment?’ (Carson, 2009). Furthermore, the following thesis will add to the existing literature by testing the causal relationship between deforestation and CO2 emissions. This is relevant since Tanzania is a developing country,

where mainly biomass is used for energy consumption and agriculture is the main source of income for a significant part of the population.

The following two research questions result from the introduction above:

1.) What are the key determinants of CO2 emissions? (Using expanded Kaya identity)

1_{The Paris agreement was signed in 2015 by 195 countries with the objective to keep the increase of global}

(6)

2.) What is the causal relationship between CO2, economic and population growth and

deforestation? (Using a model based on the Environmental Kuznets Curve - EKC literature)

Time series from 1990 to 2014 for Tanzania (extracted from the FAOSTAT and the World Bank dataset) were used in this study. First, the Kaya identity was adapted and the deforestation variable included. Subsequently, a model based on the EKC literature and the ARDL bounds test for cointegration was used in order to examine the causal relationship between CO2

emissions and its drivers. A special focus was put on deforestation and its impact on CO2

emissions. The results demonstrate that deforestation is the main driver of environmental degradation in Tanzania, while GDP per capita has a weak causal impact and population has an indirect impact through deforestation. The results also demonstrate the dependency on the agricultural sector as a source for income in Tanzania.

The rest of the paper is structured as follows: Section 2 contains the literature review, which offers a scientific framework for the relevant research questions, Section 3 introduces the econometric model and methodology, Section 4 presents the results, whereas they are shortly discussed in Section 5. Finally, conclusions of the study are summarized in Section 6. First, however, a short introduction to the issue of deforestation in Tanzania is provided in Section 1.1.

1.1 Deforestation causes and policies

(7)

sustainable management of forests and enhancement of forest carbon stocks’, is probably one that stands out the most and certainly receives most of the attention internationally. The initiative “creates a financial value for the carbon stored in forests by offering incentives for developing countries to reduce emissions from forested lands and invest in low-carbon paths to sustainable development” (unredd.net, 2018). REDD+, simply put, offers financial compensation to communities which engage in forest conservation, thus aiming to make forest conservation economically more attractive. In Tanzania, the initiative received a large amount of financial and expert support from the World Bank, Norway, Sweden, Denmark, the Netherlands and Finland but has been rather ineffective despite the large volumes of foreign aid (Lund et al., 2017). The advertised success of the initiative with its impressive figures is rather questionable according to Lund et al. (2017). The authors argue that behind acclaimed success the actual development is rather slow and that “remarkable coverage has been achieved by casting a wide net over ‘work in progress’”. Horning (2018) also shares this opinion by noting that REDD+’s success has been limited so far.

According to Lund et al. (2017), participatory forest management (PFM) played the central part in Tanzania’s Forest Policy prior to REDD+. The objective and legislative framework for PFM was set in Forest Act 2002, which aims to increase the involvement of local communities in forest management (Lund et al., 2017). There are two different approaches for PFM: Community based natural resource management (CBNRM), which recognizes “full ownership and management responsibility of villagers for an area of forest within their jurisdiction” and Joint Forest Management (JFM), which is “a collaborative management approach, which divides forest management responsibility and returns between government (either central or local) and forest adjacent communities” (Blomley and Ramadhani, 2006). Much trust was put into these strategies to include rural communities into forest conservation efforts. However, neither PFM nor REDD+ efforts have proven to be particularly successful. Both initiatives suffered from corruption and mismanagement, with REDD+ also having the problem of payments being not based on conservation performance, which led to displaced funds (Lund et al., 2017).

(8)

influx continues despite insufficient conservation performance. Ultimately, dependency on foreign aid in Africa increased and put pressure on government’s conservation commitments. Support of foreign organizations was mainly focused on the improvement of institutions, this however does not automatically result in commitment to environmental conservation. With all the conservation efforts and large inflow of foreign donor’s money, the question arises why deforestation still persists. In Africa, there is a heavy dependence of village communities on forest for livelihood as well as for spiritual reasons (Horning, 2018).

In Tanzania, communities are heavily dependent on the forests for fuelwood and charcoal collection. As reported by NAFORMA (2015), most of respondents of the study conducted by the organization in Tanzania do not see alternatives to fuelwood that are affordable. Furthermore, forested area is rather considered as a barrier for agricultural development, which ensures livelihood for many communities living in the rural areas. According to Horning (2018), in case a community is better off when exploiting the forests economically, reinforcement of rules is especially difficult since state rules are considered as being unnecessary according. Therefore, community rules are more effective in conservation efforts and mutual reinforcement of aligned interests between the local communities and state is essential. Consequently, tightening of forest legislation alone does not automatically translate into more conservation, local norms and rules need to be aligned with formal legislation (Horning, 2018).

2 Literature review

A significant amount of environmental economics academic literature exists on the concept of the Environmental Kuznets Curve and the decomposition analysis. However, there is limited literature on the integrated view even though the concepts are interrelated and help gain insight on the relationship between CO2 emissions, economic development and other

potential driving forces of environmental degradation.

2.1 The environmental Kuznets curve (EKC)

(9)

deteriorates when economic development is in its early stage but eventually reaches a turning point and improves as an economy develops (Burnett, Bergstrom and Wetzstein, 2013). This approach seems intuitive; when an economy grows rapidly and is in its first pre- and industrialization phase, the natural resources are being used intensively, which results in higher emissions. The assumption is that countries with low incomes are often too poor to invest in countermeasures and disregard environmental consequences. The underlying assumption of the EKC, thus, is that people care more for the environment when they reach a certain standard of living. In later stages of economic development and change towards service and information intensive industries, institutions may become more effective, technologies cleaner or more efficient and people tend to value the environment more (Dinda, 2004). The described relationship between environmental degradation and income per capita during the different phases of economic development is illustrated in Figure 1.

Figure 1 - Environmental Kuznets Curve

Source: Rashid Gill, Viswanathan and Hassan (2018)

(10)

(for most pollutants the turning point is US$8,000 per capita) is reached, air and water quality appear to benefit from economic growth while in very poor countries economic growth may result in environmental degradation. However, the authors also remark that there is no reason to believe that this process happens automatically. Harbaugh, Levinson and Wilson (2002) on the other hand, find little empirical support for the EKC relationship between several air pollutants and national income using the data previously studied by Grossman and Krueger (1995) but more extensive and corrected. The authors additionally argue that the empirical evidence provided in other studies in favor of the hypothesis is less robust than appears and is sensitive to data variations and econometric specifications (Harbaugh, Levinson and Wilson, 2002; Rashid Gill, Viswanathan and Hassan, 2018). Churchill et al. (2018) test the ECK hypothesis using panel data for 20 OECD countries in the period of 1870 to 2014 as a response to the critique of the EKC literature which often used short time series in the past. Short time series might result in a lack of adequate overlapping growth periods over time across countries and might be the reason why the empirical results on the EKC are lacking consensus. The authors find evidence in favour of the EKC hypothesis with turning points for individual countries varying from US$5,936 to US$129,58. They conclude that of the nine countries with EKC, three countries show an N-shaped relationship, one an inverted N-relationship and five have the traditional inverted U-shaped relationship. However, the authors also remark CO2

emissions will still be rising for some time in most countries and actions for CO2 reduction and

increased renewable energy consumption are indispensable (Churchill et al., 2018).

Separate studies used different econometric approaches (Burnett, Bergstrom and Wetzstein, 2013) but there are some common characteristics regarding the data as well as the methods. From a methodological point of view, the studies reviewed by Shahbaz and Sinha (2016) used cross-sectional panel data with panel regression tests and time series data for single countries, where ARDL bounds test has been mostly used. Different environmental quality indicators such as air, water quality and other environmental quality indicators have been used due to the absence of a single environmental indicator. Regarding the turning point, it varies for different pollutants or environmental indicators with the estimated turning point for most of the pollution indicators being in the range of US$3,000-10,000 (Dinda, 2004). After using a meta-analysis, which synthesizes results of similar empirical studies for the review, Dinda (2004) identified that significant EKC is likely to be found for short-term and local pollutants like SO2, SPM, NOx and CO rather than for the global indicators such as CO2. The following

generalized model was used in most of the studies: 𝐶_𝑖𝑡 = 𝛼_𝑖 + 𝛽₁𝑌_𝑖𝑡+ 𝛽₂𝑌_𝑖𝑡2_{+ 𝛽}

(11)

where C is CO2 emissions, Y is economic growth, D is the additional explanatory variable

which is context specific, i are the cross sections, t are the time series, 𝛼 is the constant term,

ß is the coefficient and 𝜖 the standard error term (Shahbaz and Sinha, 2016). The studies have

been shifting focus from fossil fuels to renewable energy (e.g. Jebli, Youssef and Ozturk, 2015) over the years due to an energy consumption pattern transformation. Apart from energy consumption (e.g. Acaravci and Ozturk, 2010) a variety of control variables have been used including macroeconomic indicators such as financial development (e.g. Destek and Sarkodie, 2019), trade (e.g. Raza and Shah, 2018) and population growth (e.g. Sohag et al., 2017).

2.1.1 Critique of the environmental Kuznets curve

As demonstrated by different empirical studies, there is some empirical evidence for the inverted U-shaped relationship but there is also a lot of controversy (Dinda, 2004). According to several economists, supporting evidence is fragile to data modifications and a clear causal income-pollution relationship has not been revealed yet (Carson, 2009; Harbaugh, Levinson and Wilson, 2002). The main critique point of the EKC is the assumption that economic growth has a causal relationship with environmental degradation and the automatic improvement of the latter. Another critique is that there are several econometric problems while interpreting the EKC estimates such as the problem of integrated variables, omitted variables and identification of time effects. Tests for integrated variable typically find that GDP per capita and CO2

(12)

(13)

costs are enormous (Rashid Gill, Viswanathan and Hassan, 2018). Furthermore, the role of population and technology has been mostly disregarded in the EKC studies.

To sum up, based on the mixed evidence and critique mentioned above, there is room for improvement. For example studies may put more emphasis on the identification of dominant factors explaining the EKC relationship if there is any (Dinda, 2004). More information on the dominant drivers can be provided by a decomposition analysis that considers factors such as technology and population.

2.2 The Kaya identity

In a decomposition analysis the emissions are broken down into the proximate sources of emissions (Stern, 2017). Ehrlich and Holdren (1971) first introduced a pollution decomposition equation long before the EKC debate began; the equation was later expanded and called IPAT, which relates Impact (e.g. pollution) to Population, Affluence (e.g. per capita income) and Technology (Carson, 2009). The equation is also known in a different form as the Kaya identity, where the total CO2 emissions are a product of population, per capita GDP, energy

use per capita and CO2 emissions per unit of energy consumed. This form of equation is used

in the Intergovernmental Panel on Climate Change (IPCC) estimations of future CO2 emissions

(Carson, 2009). 𝐶𝑂2 = 𝑃𝑂𝑃 × 𝐺𝐷𝑃 𝑃𝑂𝑃× 𝐸𝑛𝑒𝑟𝑔𝑦 𝐺𝐷𝑃 × 𝐶𝑂₂ 𝐸𝑛𝑒𝑟𝑔𝑦 (2)

The Kaya identity as in equation (2) includes four relevant kinds of driving forces: [CO2/Energy] describes carbon intensity of energy consumption, [Energy/GDP] describes energy intensity which is equal to energy per unit of GDP, [GDP/POP] describes GDP per capita and population is defined by [POP].

The studies conducting a decomposition analysis argue that emissions can be mainly reduced through time-related technique effects specifically directed at the emissions reduction (Stern, 2017). As noted by Carson (2009) this view is entirely different to the EKC, since it assumes that population growth together with growing affluence are the driving forces of environmental degradation while, as noted above, the EKC assumes that at a certain point economic growth leads to a decrease in negative effects on the environment. The EKC assumption can be supported by the argument that technological progress can be resource conserving and have a positive pollution reducing impact, while also accelerating at a higher rate than population and income.

2.3 Modelling approaches to tropical deforestation

(14)

will be provided in this section. Interestingly, most of the selected studies mentioned in the literature review focused either on economic drivers of pollution or deforestation. However, to the best of my knowledge, deforestation as one of the control variables next to the economic factors has not been considered in this context. This study is trying to explain the causal relationship between deforestation and CO2 emissions while also considering the economic

drivers behind this relationship. Therefore, an insight into the modelling approaches to tropical deforestation is necessary in order to attain an integrated view and to identify the explanatory variables used in the literature for this phenomenon.

(15)

deforestation. In these models input and output prices are hypothesized to have an impact on the decisions of households. The last approach presented is that of institutional models, in which institutional factors are the explanatory variables behind tropical deforestation (Barbier, 2001). Institutional structure and policy significantly affect the tropical deforestation process since they shape the interconnection between other driving factors such as population growth, agricultural decisions and commercial logging on the one side and deforestation on the other (Bhattarai and Hammig, 2001).

2.4 Research Framework

Several authors analyzed the role of energy consumption in CO2 emissions using the EKC

framework. This approach may, however, be inappropriate for a developing country such as Tanzania. In Tanzania a high proportion of population does not have access to electricity and main income source is the agricultural sector, which is less fossil fuels intensive than the transportation and manufacturing sectors. As shown in Table 1 total CO2 emission including

Land-Use, Land-Use Change and Forestry (LULUCF) are considerably higher than when excluding this sector, which demonstrates that the LULUCF sector is the main driver of CO2

emissions in Tanzania. According to IPCC, land use change is often associated with change in land cover. This can be due to forest clearing for conversion of land to agriculture. Forests act as vast carbon sinks which when destroyed release CO2 through for example biomass

combustion (Ipcc.ch, 2018). Consequently, deforestation might be a major driving force for CO2 emissions increase. Since forests act as an important source of income to local

communities in the developing world and are one of the richest biological systems offering habitats for a diversity of flora and fauna (Chiu, 2012) the issue of deforestation is even more alarming. In Tanzania forest area decreased by 9,488 (1,000ha) between 1990 and 2014, which accounts for a decrease of 17% (FAOSTAT, 2018). This is why deforestation will be introduced as one of the explanatory variables in the following analysis.

Table 1 - Total CO2 Emissions

Year Total CO2 (excluding Land-Use Change and Forestry) (MtCO2)

Total CO2 (including Land-Use Change and Forestry) (MtCO2)

(16)

1997 2.94 224.11 1998 2.77 234.12 1999 2.59 230.16 2000 3.02 213.61 2001 3.21 232.96 2002 3.66 217.64 2003 3.88 233.64 2004 5.22 233.43 2005 5.75 243.95 2006 6.08 222.17 2007 6.01 232.30 2008 6.21 231.75 2009 6.07 225.64 2010 7.28 226.96 2011 9.04 214.79 2012 11.22 215.51 2013 11.45 210.66 2014 11.77 210.69

Source: FAO 2016, FAOSTAT Emissions Database

(17)

the explanatory variables based on the EKC hypothesis, while population is included based on the Kaya Equation (2). Additionally, deforestation is included in the analysis. Furthermore, the impact of population growth and economic prosperity on deforestation will be analyzed since these are some of the most cited factors influencing forest clearing as mentioned above in Section 2.3.

Figure 2 - Conceptual framework

3 Data, Model Specification and Methodology

Following the academic literature it is appropriate to use the ARDL methodology for time series with a small sample size. The choice of the ARDL methodology is due to its increased flexibility that is offered by a consideration of different lag lengths (Shahbaz and Sinha, 2018). The methodology will be explained in more detail in Section 3.2. To the best of my knowledge the impact of deforestation as an explanatory variable on CO2 emissions has not been

considered in the context of the EKC models so far. It might, however, offer a better insight in what drives the emissions in the context of Tanzania.

3.1 Data

To examine the effect of the variables deforestation, per capita GDP and population on the CO2 emissions in Tanzania, time series from the period 1990 to 2014 have been considered.

For the CO2 emissions including the LULUCF sector and arable land, the data was extracted

from the FAOSTAT database, while for the remaining variables (GDP per capita and Population) the data was extracted from the World Bank’s World Development Indicators (WDI) online database.

The variables used in this study are GDP per capita in current $US, total population in thousands, CO2 emissions including Land Use and Forestry (measured in metric tons) extracted

(18)

the FAOSTAT2_{database. As mentioned in Section 2.3, arable land expansion for agricultural}

purposes leads to deforestation in developing countries, which simultaneously results in less CO2 being absorbed by the forests. The variables were used with their natural logarithms. This

is useful to reduce heteroskedasticity as well as to acquire the growth rate by differentiating logarithms of the variables (Acaravci and Ozturk, 2010). Appendix Table A provides the definitions of the variables and their sources while summary statistics of the variables is provided in Appendix Table B.

It is worth mentioning that there are limitations regarding the data set. First and foremost, the number of observations is rather small due to the limited period of time (1990 to 2014) observed for a single country. Secondly, institutional quality (e.g. political stability, ownership rights, corruption etc.) might be an important explanatory variable, which is omitted in this analysis due to data limitations. Data sets including the institutional factors only exist for a small number of tropical developing countries and the factors do not vary much over time (Barbier, 2001).

3.2 Methodology

In the following analysis of the dynamic causal relationship between deforestation and CO2 emissions, the ARDL bounds test is used for time series in the period between 1990 to 2014. According to Shahbaz and Sinha (2016), ARDL bounds and subsequent cointegration test is the most popular approach from the methodological point of view when using time series data for examination of the relationship between economic growth and environmental degradation. Since forests do not only provide new agricultural land but are also a source of income for the poor population, deforestation may also have impact on real GDP (Chiu, 2012). At the same time GDP and population increase may have impact on deforestation. This is where an ARDL model offers a great amount of flexibility and an opportunity to test the causal relationship among all the variables.

3.2.1 ARDL methodology

Autoregressive distributed lag or ARDL models are appropriate for testing a dynamic relationship between the variables where time series observations are correlated. In general there is a distributed effect if economic decision or action taken at one point in time t has implications on the economy at times t+1, t+2 etc. (Hill et al. 2012). The model contains the lagged values of the dependent variables as well as the current and lagged values of the explanatory variables. In other words, the implications from the change of explanatory variable

(19)

are distributed over a certain number of periods of time. Regarding the lagged value of the dependent variable, current rate of pollution for example depends on the levels of pollution in the previous period, meaning that the dependent variable is positively correlated with its value lagged one period. Furthermore, different lag lengths can be assigned. To set lags for the explanatory variables is important in this context because some independent variables such as deforestation and economic growth might not have an immediate effect but rather a lagged one.

In general, the ARDL models are better suitable for a small number of observations exactly for the reason of having the possibility of using the lagged values of the explanatory as well as the dependent variable. The approach of the ARDL bounds testing proposed by Pesaran et al. (2001) makes testing of the relationship between variables possible independently of if they are purely I(0), purely I(1) or mutually cointegrated, which is another advantage of this approach. Time series are said to be integrated of order one I(1) if they become stationary after taking the first difference while they are integrated of order zero I(0) if they are stationary. The clear advantages of the method are that some of the variables may be stationary at I(0) and some not, it also gives the possibility of cointegration among the different variables. The mean and variance are constant over time if the time series are stationary, meaning that they are not explosive nor trending. There is a general rule that stationary (time series with a non-constant mean) time series variables should not be used since the regressions are said to be spurious in that case. Spurious regressions can show significant results using non-stationary time series even though the data is unrelated (Hill et al. 2012).

To conclude, the ARDL approach has several advantages. As mentioned before, the lagged values of explanatory as well as dependent variable, a mixture of I(0) and I(1) data can be used and by using a sufficient number of lags, serial correlation of errors can be avoided (Hill et al., 2012; Pesaran et al., 2001). Furthermore, each of the variables can be used as dependent variable and the ARDL approach corrects for endogeneity of independent variables because there is no residual correlation. Therefore, it provides unbiased estimates and valid t-statistic even if there are endogenous regressors (Nkoro and Uko, 2016; Sugiawan and Managi, 2016). According to Sugiwan and Managi (2016) the ARDL approach is also appropriate for a small amount of observations, which reassures its use in this study since the time period observed is rather short (1990 to 2014). Following, the individual steps of the ARDL method will be listed.

(20)

any of the variables are integrated of order two. The ADF test is described in more detail in Section 3.2.2. The next step after ruling out the integration of the variables of order two I(2) is the performance of the ARDL bounds test for cointegration suggested by Pesaran et al. (2001) to see if there is any long-run relationship between the time series. In case of cointegration the long-run error correction model, which estimates short-run and long-run causal relationships, is estimated (Hill et al., 2012). Consequently, a series of diagnostics tests are performed to ensure the robustness of the model. According to Pesaran et al. (2001) one of the key assumptions for the ARDL bounds testing approach is that the errors are serially independent. Therefore, a Durbin Watson test for serial correlation will be performed. Durbin Watson test does not rely on a large sample approximation for the test of serial correlation of errors which makes it the choice for the following analysis (Hill et al. 2012). If the test indicates no serial correlation then it can be assumed that correlation in the errors has been eliminated by using a sufficient number of lags. The choice of a sufficient number of lags is especially important for the left hand side of the equation (Hill et al. 2012). Additionally, Breusch Godfrey test for serial correlation of the errors is conducted. White test for heteroskedasticity and Jaque Berra test for normality are also performed in the post-estimation diagnostic tests. To avoid instability caused by the parameter set, a stability test is also necessary. The cumulative sum of recursive residuals (CUSUM) tests are used to examine the stability of coefficients. If the plot of the CUSUM graph remains within the 5% significance interval, the H0 of no structural breaks cannot be rejected which indicates the stability of all the coefficients in the given regression.

The main steps of the ARDL methodology, which are considered to need more detailed description, are explained below.

3.2.2 Augmented Dickey-Fuller test: order of integration

Augmented Dickey Fuller test is a procedure to test if the variable follows a unit-root process or if it is stationary. It is necessary to test if the time series is stationary and in case it is not, it needs to be tested in which order the variable becomes stationary. For the ARDL method it is irrelevant if the variables are integrated I(0) or I(1). It is however essential that they must not be I(2). If the time series is non-stationary after the first differencing, it is integrated of order higher than one. In case the variables are I(2) or higher, the F-statistics of the bounds test are unreliable. Therefore, it is necessary to perform a unit root test before carrying out the bound F-test for cointegration (Nkoro and Uko, 2016).

The hypotheses can be written as:

(21)

The null hypothesis is that the variable contains a unit root and is non-stationary while the alternative hypothesis is that the variable was generated by a stationary process. If the parameter takes value zero, the null-hypothesis cannot be rejected.

In case H0 that y=0 is rejected the series are considered to be stationary, which means they are integrated to their own order (Hill et al. 2012).

3.2.3 ARDL bounds test for cointegration

To overcome the problem of non-stationarity, cointegration techniques offer a powerful tool for identification of constant long run equilibrium between the variables (Nkoro and Uko, 2016). Pesaran et al. (2001) propose a bounds testing procedure for cointegration. Thereby the F-statistic for joint significance of the lagged variables is computed in order to test for long-run relationship between the variables (Sugiawan and Managi, 2016). Cointegration implies that the time series follow a common trend and they never diverge too far from each other (Hill et al., 2012). The null hypothesis of no cointegration among the variables is H0:

b1=b2=b3=b4=0 against the alternative hypothesis Ha: b1b2b3b40.

3.2.4 Error Correction Model

Error correction model is a reparameterization of the dynamic relationship between cointegrated variables with long-run relations. It can be derived from ARDL through the integration of short-run adjustments and long-run equilibrium without losing long-run information (Nkoro and Uko, 2016). Therefore, it combines the long as well as short-run effects. The error correction model can be estimated if the variables are cointegrated and thus have a long-run relationship between them (Hill et al., 2012). The adjustment coefficient or the error correction term (ECT) indicates how fast the variables adjust to their long-run equilibrium. Short-run dynamics are described through the first difference in y and its regression. Finally, long-run or cointegrating relationship that holds equilibrium is also represented by the model.

3.3 Formulation of the model

3.3.1 Kaya Decomposition - estimated effect of Deforestation

In Table 2, CO2 emissions for Tanzania excluding LULUCF sector were decomposed

using the original Kaya decomposition calculation in equation (2) for the years between 1990 and 2014. On the one hand, the results demonstrate that GDP per capita growth was the main driving force behind the increase in CO2 emissions excluding LULUCF while population

(22)

Table 2 - Decomposition of CO2 emissions using the original decomposition equation

Absolute Difference Proportion (%)

CO2 intensity 197 148.84

Population 204 154.00

GDP per capita 438 330.80

Energy intensity -706 -533.64

Total3 ₁₃₂ _100.00

Source: own calculations based on WDI(2018)

The decomposition in Table 2, however, only explains a small fraction of the CO2

emissions in Tanzania since it does not include the LULUCF sector, which as shown in Table 1 is the main driver of emissions in Tanzania. Therefore, a new equation (3) for the decomposition of CO2 emissions including LULUCF is proposed, where deforestation is one

of the components instead of energy use.

𝐶𝑂₂ = 𝑃𝑂𝑃 ×𝐺𝐷𝑃 𝑃𝑂𝑃 × 𝐷𝐹 𝐺𝐷𝑃× 𝐶𝑂₂ 𝐷𝐹 (3)

where CO2 are the Total CO2 emissions including Land-Use Change and Forestry

(LULUCF) (MtCO2), POP is population, DF stands for deforestation, for which arable area is used as a proxy. The decomposition in equation (3) tries to identify how much the change of components population, GDP per capita, deforestation intensity and carbon intensity of deforestation contribute to increases in CO2 emissions. In Figure 3 the decomposition of

equation (3) is illustrated with the base year 1990.

Figure 3 - Decomposition using Deforestation variable

Source: own calculations based on FAOSTAT(2018) and WDI(2018)

The results in Figure 3 demonstrate that the total CO2 emissions including the LULUCF

sector were rather stable in the observed period from 1990 to 2014 while GDP increased almost sixth fold in the years between 1990 and 2014. Contrarily, when CO2 emissions excluding the

3_{Functional form: CO2 intensity*Population*GDP per capita*Energy intensity, where CO2 intensity is measured} in kg per kg of oil equivalent energy use, Energy intensity is measured by kg of oil equivalent energy use per capita, Population is measured in thousands and GDP per capita is measured in current $US. Energy use is defined as use of primary energy before transformation to other end-use fuels (WDI, 2018). All data was extracted from the World Band dataset.

0 20 40 60 19 90 19 91 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 20 08 20 09 20 10 20 11 20 12 20 13 20 14

Decomposition

Population (People) GDP per Capita DF/GDP CO2/DF

(23)

LULUCF sector are considered the trend was upwards (see Table 1). Therefore, it can be concluded that at the moment the total CO2 emissions are driven by the LULUCF sector, where

the rest represents only a neglectable fraction. However, this fraction increased in the observed time period considerably. With GDP per capita as well as population growing, CO2 emissions

excluding the LULUCF sector are expected to grow significantly in the future according to the results in Table 2. If we assume that deforestation will also persist in the future, net sink of carbon in the forests will decrease and more CO2 emissions will be released through the burning

of charcoal, which will lead to an increase of the total CO2 emissions in Tanzania.

These results indicate that GDP per capita and increase in Population have stronger effect on CO2 emissions from other sectors than Land Use and Forestry (e.g. industrial process,

energy). Consequently, it can be concluded that Kaya decomposition does not appropriately explain CO2 emissions in Tanzania where the LULUCF sector is prevailing. Figure 3

demonstrates that deforestation intensity, total CO2 emissions as well as CO2 intensity of

deforestation are rather constant. However, it is difficult to conclude based on Figure 3 if there is any causal relationship or long-run convergence between the variables. Therefore, we will proceed with the estimation of the model to test for level relationship between the variables.

3.3.2 Model based on EKC literature

As mentioned above the ARDL bounds testing approach for cointegration by Pesaran et al. (2001) is used to test for a long-run or cointegrating relationship that holds equilibrium between income, population growth, deforestation and environmental degradation. In this case environmental degradation is represented by CO2 emissions. The model is based on the

unrestricted error-correction model by Sugiawan and Managi (2016) where the authors tested the environmental Kuznets curve in Indonesia.

∆𝑙𝑛𝐶𝑂₂ = 𝛽₀+ ∑ 𝛽₁ 𝑝 𝑖=1 ∆𝑙𝑛𝐶𝑂₂_𝑡−1 + ∑ 𝛽₂ 𝑞 𝑖=0 Δ ln 𝐺𝐷𝑃_𝑡−𝑖 + ∑ 𝛽₃ 𝑟 𝑖=0 Δ𝑙𝑛𝑃𝑂𝑃_𝑡−𝑖 + ∑ 𝛽4𝛥𝑙𝑛𝐷𝐹𝑡−𝑖 𝑠 𝑖=1 + 𝜆1𝑙𝑛𝐶𝑂_2𝑡−1+ 𝜆2𝑙𝑛𝐺𝐷𝑃𝑡−1 + 𝜆₃𝑙𝑛𝐷𝐹_𝑡−2+ 𝜆₄𝑙𝑛𝑃𝑂𝑃 + 𝜀 (4)

where  is the short-run coefficient and  is the long-run multiplier. The variable Carbon dioxide emissions (CO2) is used as dependent variable to examine the intensity of

environmental degradation. The level of carbon emission is measured by taking log of annual CO2 emissions including land-use change and forestry measured in metric tons. Following

(24)

GDP per capita is used in order to examine the impact of economic growth on carbon emissions. The log of real GDP per capita in current US$ is taken. Since Tanzania is a developing country, where income is rather low and substantially lower than the turning points found in the literature as described in Section 2.1, a linear relationship is assumed and the log of GDP per capita squared is left out of scope in the model used for the following analysis. Since agriculture is the dominant source of income, it is also assumed that Tanzania is in the pre-industrial economic development phase (see Figure 1). As described above, a positive relationship is expected between economic growth and carbon emissions; when economic growth increases the level of CO2 emitted also increases in the initial development phase of a

country. According to Kaya decomposition logic, growth in population results in carbon emissions increase. This is why the variable is included on the right side of the equation. To examine the impact of population, the log of population is used. Log of arable land in thousand ha is used as a proxy for the deforestation variable following the studies by Chiu (2012) and Barbier (2001). The intuition behind using arable land as a proxy is that agricultural development is the main driving factor for land expansion (Barbier, 2001; see also Section 2.3) and therefore also for deforestation. It is expected to have a positive sign since the more forests are destroyed the less they can act as a carbon dioxide sink and the more CO2 is released

through the burning of wood.

4 Results

(25)

4.1.1 Augmented Dickey-Fuller test for unit roots

H0: variable contains a unit root

Ha: variable was generated by a stationary process

Null hypothesis can be rejected for the variables GDP and Population after including the trend option. The result demonstrates that the variables are generated by a stationary process. However, null hypothesis cannot be rejected for the variables DF and CO2 at lag 1, which

indicates that the level of the variables is non-stationary. After using the first difference they become stationary indicating that the variables are integrated of order one. Therefore, the minimum number of differences required to obtain stationary time series for DF and CO2 is

one. It can be concluded that the variables are integrated of mixed orders I(1) and I(0) while non are I(2) (See Appendix Table C). The ARDL model retains the usual interpretation even when the variables are a mixture of I(1) and I(0) (Pesaran et al., 2001). Therefore, we can further proceed with the ARDL bounds test for cointegration.

4.1.2 Bounds test for cointegration

Ho: no cointegration Ha: H0 is not true

For the bounds test for cointegration after Pesaran et al. (2001) the log transformation of the time series was used and CO2 taken as the dependent variable. The results for the test show

that F statistic at 9.65 is higher than the critical values for I(1) regressors at 1% (See Appendix Table D). Therefore, the H0 of no levels relationship can be rejected. We can conclude that there is a long-run cointegration between the variables included in the model. Consequently, an error correction model (ECM) can be formulated because of a level relationship between the variables.

4.1.3 Diagnostic tests following the model estimation

4.1.3.1 Serial correlation

The assumption of serial independency of residuals is given for the estimation of the model. It indicates that the lags have been used appropriately in order to mitigate the residual correlation problem (Pesaran et al., 2001). Therefore, the Durbin-Watson and Breusch Godfrey tests for serial correlation are conducted. The Durbin-Watson test at 2.3 indicates there is no serial correlation. Breusch-Godfrey LM test for autocorrelation at 0.18 (lag 1) and 0.39 (lag 2) reassures the result. By conducting the Breusch Godfrey test, we cannot reject the H0 of no serial correlation.

4.1.3.2 Homoskedasticity, normality and stability of the model

(26)

rejected. Instability test CusumSQ graph (see Figure 4) further demonstrates that the model is dynamically stable.

Figure 4 - CUSUM squared graph

Thus, it can be concluded that the model is suitable for modeling the relationship between the time series and there is no apparent issue with the model. Therefore, we can proceed with it.

4.1.4 Interpretation of the ECT model

The error correction term indicates the speed adjustment to restore equilibrium in a dynamic model. The ECT coefficient shows how quickly variables converge to equilibrium and the coefficient should be statistically significant with a negative sign. According to Banerjee et al. (1998) a highly significant ECT coefficient confirms that there is a stable long-run relationship. As can be seen from the results in Table 3, the coefficient is highly significant at 1% and negative. Therefore, the hypothesis of long-run relationship can be confirmed. The coefficient implies that a deviation from long term growth rate in CO2 rate is corrected by

126% by the following year. There is some controversy regarding what are the limits for the ECT coefficient, as a result, there is no proven standard.

Therefore, equilibrium has been proven. Furthermore, variable DF has both short-run and long-run components which are significant at 5% confidence level, which indicates its strong causal effect on the dependent variable. Variable GDP, however, only demonstrates a short-run significant component at 10%, which indicates its weak causal effect. The tests demonstrate that there is no serial correlation of residuals. The model is also dynamically stable based on the results from the CUSUM graph and the null hypothesis of homoskedasticity cannot be rejected. The results demonstrate moderate support for the assumption of economic growth having strong impact on CO2 emissions. This, however, is less surprising since most of the

CO2 emissions stem from the LULUCF sector, where deforestation is expected to have major

(27)

Table 3 - Error Correction representation of ARDL model (CO2 emissions as dependent variable)

Table 3. Error Correction representation of ARDL model (1, 1, 0, 2) Dependent variable = CO2

Variables Error correction rate Sources of causation long-run Short-run

GDP 0.0762 (0.0534) Population -0.044 (0.0358) Deforestation -0.312*** (0.0939) L.CO2 -1.260*** (0.206) L.GDP 0.161* (0.0797) L.Deforestation 0.425** (0.196) LD.Deforestation 0.312 (0.194) Constant 9.984*** (1.787) R-squared 0.74 0.74 0.74

Notes: Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

4.2 Estimation of the model using Deforestation as the dependent variable

As mentioned in Section 3.2.1 one of the advantages of the ARDL methodology is the fact that the independent variables can also be used as dependent variables. Therefore, in order to explain deforestation, the model was used with deforestation as the dependent variable while leaving out CO2 emissions as control variable since there is no indication in the literature that

(28)

significant and has a negative sign, which indicates a long-run stable relationship between the variables. The ECT coefficient implies that a deviation from long term growth rate in deforestation rate is corrected by 53% by the following year. The diagnostic tests indicate that there is no serial correlation (Durbin Watson d-statistic is 2.32 and Breusch Godfrey chi squared statistic is .31 with one degree of freedom). With White’s test for heteroskedasticity at (.35), H0 of homoskedasticity cannot be rejected. However, CUSUM squared graph rejects the H0 of no structural breaks at 5% confidence level (see Figure 5).

The graph suggests a structural break in the years between 2001 and 2007. Agricultural productivity is assumed to have strong impact on deforestation rates (see Section 2.3); when agricultural productivity is low more hectares of land are needed to produce the same amount of output. A look at the data for agricultural productivity (see Table 4 below) shows that average annual growth of agricultural output (%) as well as total factor productivity (TFP) growth (%) in the years 2001 to 2005 were higher than in the period before (1991-2000) as well as in the period after (2006-2011) 4_{. Table 4 demonstrates that agricultural productivity}

did not show a consistent upward trend, which might have resulted in the structural break.

Table 4 - Agricultural productivity

1991-2000 2001-2005 2006-2011

Output growth (%) 1.8 5.8 4.8

TFP growth (%) -0.6 2.8 2.1

Source: IFPRI (2013)

Furthermore, the R squared is lower (.41) than in the original model. A short insight in the literature in Section 2.3 makes evident that there are certainly more drivers of deforestation

4 _{International Food Policy Research Institute defines TFP as “the ratio of total output (crop and livestock} products) to total production inputs (land, labor, capital and materials). An increase in TFP implies that more output is being produced from a constant amount of resources used in the production process.” Output is the gross agricultural output measured in constant 2004-2006 $US (IFPRI, 2013).

(29)

than population and a substantial extension of the model would be necessary in order to explain deforestation. Among those drivers are weak institutions, poverty, agricultural expansion, poor technological advances etc. (Horning, 2018). Since the study focuses on the dynamics between the CO2 emissions and deforestation, these drivers are out of the scope of this study.

The result might indicate the problem of omitted variables since population and population density are only two of the causes leading to deforestation. As mentioned above illegal timber trade, increased agriculture as well as excessive biomass consumption are the most obvious reasons for deforestation. Furthermore, institutional quality and the ability to enforce legislation also influence increase in deforestation.

Table 5 - Error Correction representation of ARDL model (Deforestation as dependent variable)

Table 5. Error Correction representation of ARDL model (1, 1, 2) Dependent variable = Deforestation

Variables Error correction rate Sources of causation long-run Short-run

GDP 0.247 (0.153) Population 0.756* (0.367) L.Deforestation -0.527** (0.181) L.GDP 0.175 (0.100) L.Population 7.216*** (2.229) LD.Population -2.761*** (0.930) Constant 2.671** (1.062) R-squared 0.574 0.574 0.574

4.3 Estimation of the model using Income as the dependent variable

(30)

and agriculture is the primary income source in Tanzania. In this context, the positive and highly significant sign of the Deforestation variable (which for arable land was used as a proxy) in Table 6 makes intuitively sense. It demonstrates that increase in arable land leads to increase in GDP per capita. This also explains why deforestation still persists in Tanzania: clearing of the forests is economically more valuable than its conservation for rural communities. Income can be generated through clearing of the forests for agricultural purposes as well as for (il-)legal timber trade. The results in Table 6 also show that population increase has a positive impact on economic growth. Both long term parameters have a positive sign and are significant. The adjustment rate of (.48) is significant and has a negative sign, which once again indicates a long-run stable relationship between the variables. )

Table 6 - Error Correction representation of ARDL model (GDP per capita as dependent variable) Table 6. Error Correction representation of ARDL model (2, 0, 0)

Dependent variable = GDP per capita

Variables GDP L1. GDP 0.894*** (0.143) L2. GDP 0.378** (0.134) Population 0.316*** (0.0524) Deforestation 0.672*** (0.167) Constant -4.088*** (1.207) R-squared 0.993

5 Discussion

The model in Equation (4) was tested using CO2 emissions, deforestation as well as GDP

per capita growth as the dependent variables. The results demonstrate that there is a long-run relationship between the variables. As discussed in the previous section deforestation has a strong causal relationship on CO2 emissions in Tanzania, while population is insignificant,

which is against the assumption made by the Kaya decomposition. The results also demonstrate that population growth is the main driver of deforestation. Therefore, population has a rather indirect effect on CO2 emissions including the LULUCF sector through its impact on

(31)

demonstrated in Table 5, GDP per capita does not show any significance. This might be due to the fact that population growth exerts more pressure on the forests since there is also more demand for agriculture. The model using GDP per capita as the dependent variable demonstrates that economic growth in Tanzania is dependent on deforestation as well as on population growth. This results emphasize the existing conflict between the economic viability of deforestation, the pressure on the resources due to population increase and its negative impact on the environment. Furthermore, the analysis performed in this study indicates that conventional methods to predict CO2 emissions like the Kaya decomposition might be too

generalized in case of developing nations with a vast areas of land covered by forests. The strong academic focus on fossil fuels, GDP per capita and population growth as the drivers of environmental degradation results in a neglection of the problem of deforestation and its impact on CO2 emissions. This is problematic since deforestation does not only lead to an increase in

CO2 emissions but also to a loss of biodiversity.

The main contribution of this study is the shifted focus from the fossil fuels and industrialized economies to a developing country and the issue of deforestation, which has not received much attention in the literature yet. This is relevant, since many countries in the developing world experience fast economic growth while also being in the first stages of their development. Therefore, there is still potential to pursue a sustainable development path. The main limitation of the thesis is the limited amount of data available. The time period in the analysis is rather short (1990-2014) and only considers deforestation after it received increased attention in Tanzania and internationally. It would be interesting to observe the deforestation rates before this period to see if the newly won attention at least decreased the degree of deforestation to some extent. The literature as well as persistent deforestation rates suggest that conservational efforts have not proven to be effective yet. Further examination and detailed structural decomposition of CO2 emissions using input-output data would provide more insight

on the exact composition of CO2 emissions in Tanzania. The detailed proportional impact of

(32)

6 Conclusion

Recently, the driving forces of environmental degradation and increase in CO2 emissions

received much attention in academia. Modelling of implications of economic development on the environment gained momentum, where the EKC and decomposition analysis are the two prominent streams in the literature. It is suggested by the two concepts that there is a causal relationship between economic growth and population growth on the one side and environmental degradation on the other. This study, however, suggests that the issue of tropical deforestation is the dominant driving force of environmental degradation in Tanzania next to the economic and demographic factors.

By employing the ARDL bounds test for cointegration, the study assessed the dynamic causal relationship between CO2 emissions and deforestation as well as economic growth and

population growth. The data from the World Bank and FAOSTAT databases was used for this purpose and the period from 1990 to 2014 was examined. First, the CO2 emissions excluding

the LULUCF sector were decomposed using the Kaya identity. The decomposition results indicate that GDP per capita and population growth are the main contributors to the increase in CO2 emissions. Since a substantial part of the emissions is stemming from the LULUCF

sector in Tanzania, the Kaya identity was adapted using deforestation as one of the explanatory components instead of the conventional energy use. The results demonstrate that the original Kaya identity does not appropriately depicture the drivers of CO2 emissions in Tanzania.

Therefore, a model based on the EKC literature and the ARDL bounds test for cointegration was used in order to examine the long-run relationship between the variables. The results confirm the assumption of deforestation having a strong impact on CO2 emissions in Tanzania.

Consistent with the EKC and the Kaya decomposition logic the results indicate that there is a (weak) causal relationship between economic growth and CO2 emissions. The EKC suggests

that especially in the initial development phase in which the agricultural sector prevails, economic growth results in higher deforestation rates (if deforestation is used as a proxy for environmental degradation). Thus, it is surprising that GDP per capita does not show significance in the estimation of the model using deforestation as the dependent variable. Demographics is cited as one of the most dominant drivers of deforestation since growing population increases the pressure on forests and leads to its depletion. This hypothesis is confirmed with the results in Table 5. Population did not show any significant results in the estimation using CO2 emissions as the dependent variable, which leads to the conclusion that

population growth has a rather indirect causal effect on the increase in CO2 emissions including

(33)

conservation of the forests. Consequently, economic growth as well as deforestation result in environmental degradation in Tanzania.

Based on the results of this study, it can be concluded that with further population growth, the deforestation rates are unlikely to decrease, unless agricultural productivity will improve and alternatives to biomass will be made available. At the same time Tanzania will industrialize its economy in the future since it is a fast growing developing country and more people are expected to get access to electricity. Therefore, CO2 emissions are likely to increase due to

deforestation as well as due to other driving forces.

(34)

References:

Acaravci, A. and Ozturk, I. (2010) ‘On the relationship between energy consumption, CO2emissions and economic growth in Europe’, Energy. Elsevier Ltd, 35(12), pp. 5412–5420. doi: 10.1016/j.energy.2010.07.009.

Ahlborg, H. and Hammar, L. (2014) “Drivers and Barriers to Rural Electrification in Tanzania and Mozambique - Grid-Extension, Off-Grid, and Renewable Energy Technologies,” Renewable Energy, 61, pp. 117–124. doi: 10.1016/j.renene.2012.09.057.

Apergis, N. and Payne, J. E. (2011) “Renewable and Non-Renewable Electricity Consumption-Growth Nexus: Evidence from Emerging Market Economies,” Applied Energy, 88(12), pp. 5226–5230. doi: 10.1016/j.apenergy.2011.06.041.

Barbier, E. B. (2001) ‘The Economics of Tropical Deforestation and Land Use: An Introduction to the Special Issue’, Land Economics, 77(2), p. 155. doi: 10.2307/3147087.

Banerjee, A. Dolado, J., and Mestre, R. (1998) ‘Error-Correction Mechanism Tests for Cointegration in Single Equation Framework’, Universidad Carlos III de Madrid, Open Access publications from Universidad Carlos III de Madrid. doi: 10.1111/1467-9892.00091.

Bhattarai, M. and Hammig, M. (2001) ‘Institutions and the Environmental Kuznets Curve for Deforestation: A Crosscountry Analysis for Latin America, Africa and Asia’, World Development, 29(6), pp. 995–1010. doi: 10.1016/S0305-750X(01)00019-5.

Blomley, T. and Ramadhani, H. (2006) ‘Going to Scale with Participatory Forest Management: Early Lessons from Tanzania 1’, International Forestry Review. Commonwealth Forestry Association, 8(1), pp. 93–100. doi: 10.1505/ifor.8.1.93.

Burnett, J. W., Bergstrom, J. C. and Wetzstein, M. E. (2013) ‘Carbon dioxide emissions and economic growth in the U.S.’, Journal of Policy Modeling. The Society for Policy Modeling, 35(6), pp. 1014–1028. doi: 10.1016/j.jpolmod.2013.08.001.

CAIT Climate Data Explorer. 2017. Washington, DC: World Resources Institute. Available online at: http://cait.wri.org

Carson, R. (2009). The Environmental Kuznets Curve: Seeking Empirical Regularity and Theoretical Structure. Review of Environmental Economics and Policy, 4(1), pp.3-23.

Chiu, Y. Bin (2012) ‘Deforestation and the Environmental Kuznets Curve in Developing Countries: A Panel Smooth Transition Regression Approach’, Canadian Journal of Agricultural Economics, 60(2), pp. 177–194. doi: 10.1111/j.1744-7976.2012.01251.x.

(35)

Cia.gov. (2018). The World Factbook — Central Intelligence Agency. [online] Available at: https://www.cia.gov/library/publications/the-world-factbook/geos/tz.html [Accessed 27 Sep. 2018].

Dasgupta, S., Laplante, B., Wang, H. and Wheeler, D. (2002) ‘Confronting the environmental Kuznets curve. Journal of economic perspectives’, 16(1), pp.147-168.

Destek, M. A. and Sarkodie, S. A. (2019) 'Investigation of Environmental Kuznets Curve for Ecological Footprint: The Role of Energy and Financial Development,' Science of The Total Environment, 650, pp. 2483–2489. doi: 10.1016/j.scitotenv.2018.10.017.

Dinda, S. (2004) “Environmental Kuznets Curve Hypothesis: A Survey,” ECOLOGICAL ECONOMICS, 49(4), pp. 431–455.

Ehrlich R., and Holdren, J. (1971) “Impact of Population Growth.” Science (New York, N.Y.). 171. 1212-7. 10.1126/science.171.3977.1212.

FAO (2015) 'Global Forest Resources Assessment 2015', FAO Forestry Paper No. 1. The Food and Agricultural Organization of the United Nations. Rome.

Food and Agriculture Organization of the United Nations (FAO). 2016. FAOSTAT Emissions Database. Rome, Italy: FAO. Available at: http://www.fao.org/faostat/en/#data

Grossman, G. and Krueger, A. (1995). ‘Economic Growth and the Environment.’ The Quarterly Journal of Economics, 110(2), pp.353-377.

Harbaugh, W., Levinson, A. and Wilson, D. (2002). ‘Reexamining the Empirical Evidence for an Environmental Kuznets Curve.’ Review of Economics and Statistics, 84(3), pp.541-551. Hashem Pesaran, M., Shin, Y. and Smith, R. J. (2001) ‘BOUNDS TESTING APPROACHES TO THE ANALYSIS OF LEVEL RELATIONSHIPS’, JOURNAL OF APPLIED ECONOMETRICS J. Appl. Econ, 16, pp. 289–326. doi: 10.1002/jae.616.

Hill, R. C, Griffiths, W. E, Lim, G. C. (2012) Principles of Econometrics, 4th edition, John Wiley & Sons Inc., New York, USA

Horning, N. R. (2018) The politics of deforestation in africa : madagascar, tanzania, and uganda. Cham, Switzerland: Palgrave Macmillan. doi: 10.1007/978-3-319-76828-1.

International Food Policy Research Institute (IFPRI) (2013) ‘Total and Partial Factor Productivity in Developing Countries’, https://hdl.handle.net/1902.1/20518, Harvard Dataverse, V6

(36)

Ipcc.ch. (2018). Land Use, Land-Use Change and Forestry. [online] Available at: http://www.ipcc.ch/ipccreports/sres/land_use/index.php?idp=33 [Accessed 7 Nov. 2018].

IRENA (2017), Renewables Readiness Assessment: United Republic of Tanzania, International Renewable Energy Agency, Abu Dhabi.

Jebli, M. Ben, Youssef, S. Ben and Ozturk, I. (2015) ‘The Role of Renewable Energy Consumption and Trade : Environmental Kuznets Curve Analysis for Sub-Saharan Africa Countries’, African Development Review, 27(3), pp. 288–300. doi: 10.1111/1467-8268.12147. NAFORMA National Forest Resources Monitoring and Assessment of Tanzania Mainland MAIN RESULTS MINISTRY OF NATURAL RESOURCES & TOURISM TANZANIA FOREST SERVICES (TFS) AGENCY IN COLLABORATION WITH THE GOVERNMENT OF FINLAND AND FOOD AND AGRICULTURE ORGANI (2015). Available at: http://naforma.mnrt.go.tz (Accessed: 10 December 2018).

Nkoro, E. and Uko, A. K. (2016) ‘Autoregressive Distributed Lag (ARDL) cointegration technique: application and interpretation’, Journal of Statistical and Econometric Methods, 5(4), pp. 63–91. Available at: http://www.scienpress.com/Upload/JSEM/Vol 5_4_3.pdf (Accessed: 10 November 2018).

Odhiambo, N. M. (2009) “Energy Consumption and Economic Growth Nexus in Tanzania: An Ardl Bounds Testing Approach,” Energy Policy, 37(2), pp. 617–622. doi: 10.1016/j.enpol.2008.09.077.

Raza, S. A. and Shah, N. (2018) 'Testing Environmental Kuznets Curve Hypothesis in G7 Countries: The Role of Renewable Energy Consumption and Trade,' Environmental Science and Pollution Research, 25(27), pp. 26965–26977. doi: 10.1007/s11356-018-2673-z.

Rashid Gill, A., Viswanathan, K. K. and Hassan, S. (2018) ‘The Environmental Kuznets Curve (EKC) and the environmental problem of the day’, Renewable and Sustainable Energy Reviews. Elsevier Ltd, 81(June 2017), pp. 1636–1642. doi: 10.1016/j.rser.2017.05.247.

Rotmans, J. and Swart, R. J. (1991) Modelling tropical deforestation and its consequences for global climate, Ecological Modelling. Available at: https://ac-els-cdn-com.proxy- ub.rug.nl/0304380091900383/1-s2.0-0304380091900383-main.pdf?_tid=96c80c45-4cae-4a8d-ae13-10c2eaabcc1c&acdnat=1546005443_66d6db52830730bb7880ae36a7ab7f54 (Accessed: 28 December 2018).