1 Introduction

Industrial revolution and rapid economic growth in countries have resulted into day’s much discussed phenomena like global warming and climate change. The global increasing concern for environmental sustainability is visible in a public arena; the development strategy depends largely on economic growth which causes environmental degradation. In the past two decades, Kaygusuz (2009) has indicated the increase of carbon dioxide emissions is considered one of the main causes of global warming and climatic instability. The prevalence of such problems is higher in countries such as Tunisia, where economic growth, energy security and environmental sustainability are simultaneously important. On the other hand, the use of energy natural resources as inputs in the development or production process is a problem.

Tunisia, as a developing country, economic growth represents an interesting case where it faces the difficulty to fulfill the needs of energy demand. Industries may increase pollution in developing countries such as Tunisia, due to the increased production of emissions intensity to export goods to developed countries. Currently, the economy of Tunisia is diverse economy, ranging from agriculture, fisheries, mining, manufacturing, textiles, and petroleum products, to tourism. Tunisia is not like some North African countries as Algeria and Libya which are mostly endowed with rich hydrocarbon resources. Tunisia is endowed with multi and excellent renewable resources. Thus, Tunisia should be more keen to invest in renewable energy generation projects to minimize spending and dependence on foreign energy sources. Therefore, analyses of economic and environmental impacts of concerned variables are needed to address this research gap and to improve the design and accept the energy renewable policy.

The main objective of present study is to examine the long run and the relationship between economic growth, renewable energy consumption, energy consumption and environmental pollution measured by CO2 emissions in Tunisia during the period 1990–2015. This object is achieved by the following steps: First, stationarity and Johansen co-integration are tested; second, error-correction models are estimated to test for the Granger causality between the two variables. The rest of this article is structured as follows: After the section (1) which is the introduction, this paper includes five additional sections: section (2) is a literature review; section (3) presents the data source and methodological framework; section (4) presents the empirical results and interpretation. finally, section (5) present the conclusion and policy implications.

2 Literature review

Nowadays, the causal relationship among energy, environment pollution and economic growth has been well established in empirical studies, which use diverse approaches, methods, variables and different countries and regions. However, the empirical results of these works are contradictory. Certain researchers argue that exists a bidirectional causality between energy use and economic growth; while some studies showed that exists a unidirectional relationship from energy to economic growth where the reverse. From the perspective of environmental economics, the relation among these variables depicts a matter of great importance. To the extent of our knowledge, only a single study based on the scenario of Bangladesh was conducted involving all three aspects: Energy, CO2 emissions and economic growth. On the other hand, there has had many difference between the results obtained. For example, Kraft and Kraft 1978) investigated the link between energy use and economic growth in the United States; they concluded that GNP leads to energy consumption. This seminal work was followed by an impressive number of studies on the same topic.

Several empirical literature, such as Jbir and Zouari-Ghorbel (2009), Menyah and Wolde-Rufael (2010), Pao and Yu (2011), Akpan and Akpan (2011), Alam et al. (2012), Kais and Hammami (2014), Mbarek et al. (2014) and Salahuddin and Gow (2014), Mbarek et al. (2015), Saidi and Hammami (2015a, b), investigate the relationship between energy consumption, greenhouse gas (GHG) emissions and economic growth for a number of developing and developed countries. The recent research on energy use and economic development ensures a stream of information regarding the causality sense between renewable energy and economic growth. Renewable energy consumption is inherently linked to sustainable development, emissions reductions and energy security (Reboredo 2015). Every country has different effects of renewable energy use; some countries report economic growth fetches handsome contribution in renewable energy consumption, while for some countries this source of energy contribute in economic growth. In addition, several others studies have found no relationship and/or two-way causal relationship between economic growth and renewable energy use. The summary of empirical studies regarding the causality relationship between renewable energy and economic growth are presented in Table 1.

Table 1 Summary of the studies based on energy use, renewable energy and economic growth
Full size table

We can classify the literature considered in two categories. The first category comprises studies which examine the nexus between energy and economic growth, and the second category comprises the studies that examine the relationship between renewable energy and economic growth. In this framework, Narayan and Smyth (2008), Bowden and Payne (2010), Belloumi (2009), Belke et al. (2011), Yildirim et al. (2012), Al-mulali et al. (2013), and Sebri (2015) proved the validity of the growth hypothesis implies an increase, respectively (decrease) in energy consumption leads to an increase (decrease) in real GDP. In this case, the energy box the GDP and the economy is greatly dependent on energy; Belloumi (2009), Apergis and Payne (2010a, b), Apergis and Payne (2011), Ozturk et al. (2010), Eggoh et al. (2011), Apergis and Payne (2012), Tugcu et al. (2012), Al-mulali et al. (2013), Pao and Fu (2013), Sebri (2015) proved the validity of the feedback hypothesis suggests a two-way causal relationship between energy consumption and real GDP so that an implementation of an efficient consumer policy has no negative effect on real GDP; Huang et al. (2008), Sadorsky (2009), Ozturk et al. (2010), Al-mulali et al. (2013), Pao and Fu (2013), Ocal and Aslan (2013a, b), and Sebri (2015) proved the validity of the conservation hypothesis involves that energy conservation policies resulting in reduced energy consumption have no negative effects on real GDP. This hypothesis is checked whether an increase in GDP leads to an increase in energy consumption, and Menegaki (2011), Tugcu et al. (2012), Yildirim et al. (2012) and Sebri (2015) proved the validity of the neutrality hypothesis considers that the energy consumption is only a fraction of the components of the production and its effect on real GDP is low or zero. This hypothesis is confirmed by the absence of a causal relationship between energy consumption and real GDP. We summarize the studies that based on the link between economy, energy use and CO2 emissions discussed above in the following Table 1.

3 Data source and methodological framework

3.1 Data sources

The annual data used in this the paper are taken from the World Development Indicator (WDI, 2014-CD-ROM) for Tunisia, and cover 1990–2015. The variables used are the economic growth (GDP) [proxied in GDP per capita (constant 2005 US$)], CO2 emissions (CO2) (measured in metric tons per capita), energy use (EU) (measured in kilogram (kg) of oil equivalent per capita), and Renewable energy consumption (REC) (measured in 1000 metric tons of oil equivalent). The descriptive statistics Mean, Median, Maximum, and Minimum of these variables are recorded below in Table 2.

Table 2 Descriptive statistics of variables
Full size table

A descriptive analysis of research variables shows that the average of GDP per capita in Tunisia is 8.076. The variables of renewable energy consumption are close to the average level (2.652). We check multicollinearity possible between the independent variables in our model by using the correlation matrix (Table 3). The variables of energy use and CO2 emissions are correlated positively on GDP per capita, but the renewable energy consumption is correlated with a negative manner on GDP per capita. In addition, energy use and CO2 emissions are negatively correlated to the renewable energy consumption. Finally, there is a positive correlation between CO2 emissions and energy use.

Table 3 Correlation matrix between the variables.
Full size table

3.2 Methodological framework

We examine the causality relationships short and long run between economic growth, energy use and CO2 emissions and renewable energy consumption in case of Tunisia. The analysis begins with the examination of the stationary properties of the variables by employing a battery of unit root tests, co-integration tests and Granger causality tests. For our case, the studied variables take the linear form compared by other variables as energy prices that generally take the non linear form. It for this reason we a linear model to resolve this problem. An important paper by Beckmann et al. (2014) investigated the relationship between the spot and futures prices of energy commodities from a new perspective. In fact, the movements of global commodity prices take the nonlinear form. Unlike, we have studies the dynamic links between energy consumption and economic growth that take a linear form according to several studies in this area, Menegaki (2011), Tugcu et al. (2012), Yildirim et al. (2012), Sebri (2015) and Mbarek et al. (2015).

3.2.1 Unit root tests

The most effective method to determine the order of integration of a series is based on the unit root tests. The unit root tests can detect the presence of unit root in a series. Two unit root tests are usually used, i.e. the test Augmented Dickey–Fuller (ADF) (1979) and the Phillips–Perron (PP) (1988).

  1. (a)

    Augmented Dickey–Fuller unit root tests

    It consists of testing the null hypothesis H0: ρ = 1 against the alternative hypothesis H1: ρ < 1. It is based on the least squares estimation of three models:

    $$\Delta x_{t} = \left( {\rho - 1} \right)x_{{t - 1}} + \sum\limits_{{j = 2}}^{k} {\theta _{j} } \Delta _{{t - j + 1}} + \varepsilon _{t} {\text{:}}\;{\text{Process}}\;{\text{no}}\;{\text{trend}}\;{\text{and}}\;{\text{no}}\;{\text{constant}}$$
    (1)
    $$\Delta x_{t} = \left( {\rho - 1} \right)x_{{t - 1}} + \sum\limits_{{j = 2}}^{k} {\theta _{j} } \Delta _{{t - j + 1}} + \alpha + \varepsilon _{t} {\text{:}}\;{\text{Process}}\;{\text{no}}\;{\text{trend}}\;{\text{and}}\;{\text{with}}\;{\text{constant}}$$
    (2)
    $$\Delta x_{t} = \left( {\rho - 1} \right)x_{{t - 1}} + \sum\limits_{{j = 2}}^{k} {\theta _{j} } \Delta _{{t - j + 1}} + \alpha + \beta t + \varepsilon _{t} {\text{:}}\;{\text{Process}}\;{\text{with}}\;{\text{trend}}\;{\text{and}}\;{\text{constant}}$$
    (3)
  2. (b)

    The Phillips–Perron test

    This test provides a nonparametric statistical correction of Augmented Dickey–Fuller in the presence of autocorrelation of the unknown form (AR (p), MA (q) and ARMA (p, q)). It has the following three models:

$$\Delta x_{t} = \left( {\rho - 1} \right)x_{t - 1} + \varepsilon_{t}{:}\;{\text{Process no trend and no constant}}$$
(4)
$$\Delta x_{t} = \left( {\rho - 1} \right)x_{t - 1} + \alpha + \varepsilon_{t}{:}\;{\text{Process no trend and with constant}}$$
(5)
$$\Delta x_{t} = \left( {\rho - 1} \right)x_{t - 1} + \alpha + \beta t + \varepsilon_{t}{:}\;{\text{Process with trend and constant}}$$
(6)

As in the case of the Dickey–Fuller test, the hypotheses to be tested are the same.

3.2.2 Johansen co-integration test

This is the step that follows the preliminary tests check non-stationarity of the series. The two-step approach of Engle and Granger is very restrictive. Indeed, this approach is applicable only in the case of a single co-integration relationship (so only one co-integrating vector). In addition, it poses a standardization problem; it can lead to different results depending on whether one considers the combination z t  = x t  − a − by t or \(z_{t} = y_{t} - a - bx_{t}\).

As an alternative to the approach of Engle and Granger, it instead uses the co-integration test of Johansen. This test determines the number of long-term equilibrium relationship between integrated variables same regardless of the normalization used. The Johansen co-integration test uses two statistics to determine the number of co-integrating vectors r:

  • Trace test for the existence of assumption of no co-integrating vectors given by:

    $$Trace = - T\mathop \sum \limits_{i = r + 1}^{P} ln\left( {1 - \hat{\lambda }_{i} } \right)$$

  • Testing the maximum eigenvalue for the existence of assumption exactly vectors of co-integration: \(\lambda_{max} = Tln\left( {1 - \hat{\lambda }_{r + 1} } \right) .\)

3.2.3 The error correction model

Accept co-integration, it is to accept the fact that there is a steady state relationship between the four sets of variables that have a common tendency to move in the same direction. Any momentary deviation from balance is considered random. According to Granger representation theorem, all co-integrated system implies the existence of an error correction mechanism that prevents too variables deviate from their long-term equilibrium. If cointegration allows clarify the reality and nature of discrepancies between four series theoretically linked and to model the behavior of these variables, the error correction model can explain and deduce the mechanism. In a way general, you can simply write the error correction model as follows:

$$\Delta GDP_{t} = \alpha_{1} z_{t - 1} + lagged\left( {\Delta GDP_{t} , \Delta REC_{t} ,\Delta EU_{t} ,\Delta CO2_{t} } \right) + \varepsilon_{1t}$$
(7)
$$\Delta REC_{t} = \alpha_{2} z_{t - 1} + lagged\left( {\Delta GDP_{t} ,\Delta REC_{t} ,\Delta EU_{t} ,\Delta CO2_{t} } \right) + \varepsilon_{2t}$$
(8)
$$\Delta EU_{t} = \alpha_{3} z_{t - 1} + lagged\left( {\Delta GDP_{t} ,\Delta REC_{t} ,\Delta EU_{t} ,\Delta CO2_{t} } \right) + \varepsilon_{3t}$$
(9)
$$\Delta CO2_{t} = \alpha_{4} z_{t - 1} + lagged\left( {\Delta GDP_{t} , \Delta REC_{t} ,\Delta EU_{t} ,\Delta CO2_{t} } \right) + \varepsilon_{4t}$$
(10)

where z t−1 is the term of error correction coming from the estimate of the relationship of co-integration, \(\varepsilon\) is a term of stationary error; |α 1| + |α 2| + |α 3| + |α 4| ≠ 0.

3.2.4 Causality test

Engle and Granger (1991) showed that if the variables are integrated, the classic test of Granger, based on the VAR is no longer appropriate. They recommend doing this using the error correction model. In addition, the causality test based on the vector model with error correction present the advantage of a causal relationship even if no estimated coefficient of lagged variables of interest is significant. Taking into account relations (7), (8), (9) and (10), we can rewrite Eqs. (11), (12), (13) and (14) as follows:

$$\Delta GDP_{t} = \alpha + \mathop \sum \limits_{i = 1}^{k} \theta_{1} \Delta GDP_{t - i} + \mathop \sum \limits_{i = 1}^{k} \tau_{1} \Delta REC_{t - i} + \mathop \sum \limits_{i = 1}^{k} \rho_{1} \Delta EU_{t - i} + \mathop \sum \limits_{i = 1}^{k} \sigma_{1} \Delta CO2_{t - i} + \varphi_{1} z_{t - 1} + \varepsilon_{t}$$
(11)
$$\Delta REC_{t} = \beta + \mathop \sum \limits_{i = 1}^{k} \tau_{2} \Delta REC_{t - i} + \mathop \sum \limits_{i = 1}^{k} \theta_{2} \Delta GDP_{t - i} + \mathop \sum \limits_{i = 1}^{k} \rho_{2} \Delta EU_{t - i} + \mathop \sum \limits_{i = 1}^{k} \sigma_{2} \Delta CO2_{t - i} + \varphi_{2} z_{t - 1} + \vartheta_{t}$$
(12)
$$\Delta EU_{t} = \gamma + \mathop \sum \limits_{i = 1}^{k} \rho_{3} \Delta EU_{t - i} + \mathop \sum \limits_{i = 1}^{k} \theta_{3} \Delta GDP_{t - i} + \mathop \sum \limits_{i = 1}^{k} \tau_{3} \Delta REC_{t - i} + \mathop \sum \limits_{i = 1}^{k} \sigma_{3} \Delta CO2_{t - i} + \varphi_{3} z_{t - 1} + \mu_{t}$$
(13)
$$\Delta CO2_{t} = \delta + \mathop \sum \limits_{i = 1}^{k} \sigma_{4} \Delta CO2_{t - i} + \mathop \sum \limits_{i = 1}^{k} \theta_{4} \Delta GDP_{t - i} + \mathop \sum \limits_{i = 1}^{k} \tau_{4} \Delta REC_{t - i} + \mathop \sum \limits_{i = 1}^{k} \rho_{4} \Delta EU_{t - i} + \varphi_{4} z_{t - 1} + \pi_{t}$$
(14)

By using the vector model error correction Eqs. (11), (12), (13) and (14), GDP t does not cause REC t Granger if θ 2 = φ 2 = 0; REC t does not cause GDP t if τ 1 = φ 1 = 0; GDP t does not cause EU t Granger if θ 3 = φ 3 = 0; EU t does not cause GDP t Granger if ρ 1 = φ 1 = 0; GDP t does not cause CO2 t Granger if θ 4 = φ 4 = 0; CO2 t does not cause GDP t Granger if σ 1 = φ 1 = 0; REC t does not cause EU Granger if τ 3 = φ 3 = 0; EU t does not cause REC t Granger if ρ 2 = φ 2 = 0; REC t does not cause CO2 t Granger if τ 4 = φ 4 = 0; CO2 t does not cause REC t Granger if σ 2 = φ 2 = 0; CO 2t does not cause EU t Granger if σ 3 = φ 3 = 0; EU t does not cause CO2 t Granger if ρ 4 = φ 4 = 0.

4 Empirical results and interpretation

4.1 Unit root tests

The results of the unit root tests are presented in Table 4. They show that the four series studied namely economic growth (LNPIB), consumption of renewable energy (LNREC), the use of energy (LNEU) and CO2 emissions (LNCO2) are integrated of order 1.

Table 4 Result of unit root test
Full size table

Indeed the unit root hypothesis tested in Eq. (3) with four LNGDP series LNREC, LNEU and LNCO in level is accepted at the 5% threshold. If we admit that ρ = 1 in the Eq. (3), while the β coefficient is significantly zero. We test the unit root again in Eq. (2) where β is disappeared. We accept once again the unit root hypothesis. The nullity of constant α being rejected, when it is assumed that ρ = 1 in Eq. (2), case of the consumption of renewable energy (LNREC), the use of energy (LNEU) and emissions CO2 (LNCO2), we retain those for variable model (2) to increase the Dickey–Fuller test. It is concluded that the two series are integrated of order 1 level. By performing of the Augmented Dickey–Fuller test on the series in first differences, with the same procedure, we retain the model (1) for both series. The unit root hypothesis is rejected in this case (at 1% ADF, t stat = 3925 for LGDP; at 5% ADF, t stat = −3540 for LNREC; at 1% ADF, t-stat = 11,266 to LNEU and at 1% ADF, t stat = 9926 for LNCO2).

The Phillips-Perron test applied the same strategy on the four series confirms this characterization at the 1% level with the models (4), (5) and (23). The unit root tests confirm the impossibility to reject the hypothesis that the four variables LNGDP, LNREC, LNEU LNCO2 and are integrated of order 1 (I (1)). It is, therefore, possible that they are co-integrated.

4.2 Results of co-integration test

We recall that the different sub-models tested of general model are:

  • Model 1 there are no constants and linear trends in the VAR and co-integration relationship does not include either of constant and linear trends;

  • Model 2 there is no constant and linear trend in the VAR, but the relationship of co-integration includes a constant (no linear trend);

  • Model 3 there are constant (no linear trends) in the VAR and relationship of co-integration includes a constant (not a linear trend);

  • Model 4 there are constant (no linear trends) in the VAR and the relationship of co-integration includes a linear constant;

  • Model 5 there are constants and trends in the VAR and the relationship of co-integration includes a constant and a linear trend.

Table 5 Presents the number of delays. By testing these different sub-models for different values of k, the Akaike information criterion is optimized for the models, r = 2 and k = 3. This model indicates the existence of a quadratic trend in each component of the system that is level, and then the system is written in first differences.

Table 5 Choice of number of delays k
Full size table

By testing these sub-models with a delay k = 3, we find that the optimized model is (or constant or trend), r = 2, k = 3. The Johansen test will be conducted from this model r = 2 A delay with k = 3. The test results are shown in Table 6.

Table 6 Johansen cointegration test
Full size table

This table shows that the null hypothesis None, At most 1 and At most 2 (for the test track and the maximum eigenvalue test) is rejected at the 1%.  Also, the null hypothesis At most 3 (for trace test and the test of the maximum eigenvalue) is rejected at the 5%. Both Johansen co-integration tests confirm the existence of a co-integration relationship.

4.3 Estimation of error correction model

The representation theorem of Engle and Granger, we said, demonstrates that the non-stationary series, especially those that have a unit root, must be represented in the form of error correction model if they are co-integrated, that is to say if there is a stationary linear combination between them. The estimated of the vector error correction model passes by determining the long-term relationship below, where the GDP variable is normalized:

$$LNGDP_{ - 1} = \mathop { 0.1089LNREC_{ - 1} }\limits_{{\left( {0.305} \right)}} \mathop { - 0.3211LNEU_{ - 1} }\limits_{{\left( {0.108} \right)}} \mathop { -0.2255LNCO2_{ - 1} }\limits_{{\left( {0.144} \right)}} - 0.0041$$

The estimate of error correction model is given in the table below (Table 7). The value in parenthesis represents the Student statistic associated with the coefficient estimated LNREC, LNEU and LNCO2. According to this relationship, in the long term, GDP and consumption of renewable energy, energy use and CO2 emissions go hand in hand because the GDP coefficient is positive. Thus, in the long term, an increase of 1% of GDP leads to an increase in renewable energy consumption, energy use and CO2 emissions of 1.33; 0.36 and 1.12%, respectively. The Jarque–Bera statistic indicates in fact that residues of the error correction model are normally distributed. In addition, the Zt−1 error correction term parameter is negative and significant in the equation of GDP, confirming the existence of a long-term relationship between renewable energy consumption, energy use, CO2 emissions and growth.

Table 7 Vector error correction model (VECM)
Full size table

The CO2 emissions negatively affect GDP and the impact of renewable energy consumption is positive on GDP. This implies that a 1% increase in CO2 and renewable energy decreases and increases GDP by 0.28 and 0.16%, respectively.

4.4 Granger causality test

We have seen that the existence of a co-integration relationship between these four variables resulted in the existence of a causal relationship between them in at least one direction. This causal relationship is discussed here using the Granger causality test based on vector error correction model. The results of this test are presented in the following Table 8.

Table 8 Granger causality test
Full size table

They were obtained in reality using the Wald coefficient restriction test based on each equation of error correction model. Indeed, the Wald test gives the possibility to integrate in the null hypothesis the coefficient of the term error correction. The Granger causality test reveals the existence of a unidirectional relationship running from GDP growth to the renewable energy consumption. The results are confirmed with the results of Ocal and Aslan (2013a, b) and Xu (2016). In addition, there is a unidirectional causality running from energy consumption to GDP per capita. The result also shows that there is a bidirectional relationship between CO2 emissions and GDP per capita between the renewable energy consumption and energy consumption and between CO2 emissions and energy consumption. The results are in line with the results of the Kim et al. (2010), Alam et al. (2011), Saboori and Sulaiman (2013) and Long et al. (2015). Unidirectional relationship running from CO2 emissions to renewable energy consumption is found. The result is in line with the conclusions of Farhani (2013)

5 Conclusion and policy implications

The main objective of this study is to investigate the relationship between economic growth, energy use, carbon dioxide emissions and renewable energy consumption in Tunisia using data from 1990 to 2015. Empirically, we tested the order of integration of the series (unit root tests) to ensure that they follow a random walk (only scope of Granger representation theorem). Then we tested the co-integration to determine the existence of a steady state relationship between the variables. Finally, we estimated the error correction model which aims to account in the same equation of a possible deviation from a long-term balance and short-term adjustment process of balance.

Our empirical findings of the Causality test indicate that there is a bidirectional causal relationship between energy use and CO2 emissions and a bidirectional causal relation between CO2 emissions and renewable energy consumption. Moreover, the results indicate that there is a unidirectional relationship running from energy consumption to GDP per capita and from CO2 emissions to renewable energy consumption.

One of the main tasks of energy policy is energy conservation, which means more efficient consumption of energy and a reduction of CO2 emissions through alternative energy options. In addition, the empirical results of this paper provide policymakers a better understanding of energy use–economic growth nexus; energy use-CO2 emissions nexus and renewable energy consumption-economic growth nexus to formulate energy and climate policies in Tunisia. If an increase in economic growth brings about an increase in energy consumption, the externality of energy use will set back economic growth. Under this circumstance, a conservation policy is necessary. In addition, the findings of this study have important policy implications and it shows that this issue still deserves further attention in future research.