Microsoft Word - DEM_2014_51to70.docx © 2014 Nicolaus Copernicus University. All rights reserved. http://www.dem.umk.pl/dem D Y N A M I C E C O N O M E T R I C M O D E L S DOI: http://dx.doi.org/10.12775/DEM.2014.003 Vol. 14 (2014) 51−70 Submitted October 20, 2014 ISSN Accepted December 23, 2014 1234-3862 Mariola Piłatowska, Aneta Włodarczyk, Marcin Zawada* The Environmental Kuznets Curve in Poland – Evidence from Threshold Cointegration Analysis A b s t r a c t. The article aims to look at the long-run equilibrium relationship between per capita greenhouse gas emissions and per capita real GDP (EKC hypothesis) in an asymmetric framework using the non-linear threshold cointegration and error correction methodology for Polish economy during the period 2000 to 2012 (quarterly data). To test the robustness of the results the additional explanatory variable (per capita energy consumption) is added to the EKC model. The EKC hypothesis is tested using threshold autoregressive (TAR) and momen- tum threshold autoregressive (MTAR) cointegration method. Moreover, the threshold error correction model (TECM) is implemented in order to examine both the short-run and the long-run Granger-causal relationship between per capita greenhouse gas emissions and per capita income. We found strong evidence in favour of the EKC hypothesis for the Polish case and additionally we confirmed that adjustment of deviations toward the long-run equilibrium is asymmetric. K e y w o r d s: Environmental Kuznets Curve, greenhouse gas emission, energy consump- tion, growth, threshold cointegration, Granger causality. J E L Classification: C24, Q43, Q50. * Correspondence to: Mariola Piłatowska, Nicolaus Copernicus University, Department of Econometrics and Statistics, 13A Gagarina Street, 87-100 Toruń, Poland, e-mail: mari- ola.pilatowska@umk.pl; Aneta Włodarczyk, Czestochowa University of Technology, Faculty of Management, 19B Armii Krajowej Street, 42-201 Czestochowa, Poland, e-mail: aneta.w@interia.pl; Marcin Zawada, Czestochowa University of Technology, Faculty of Management, 19B Armii Krajowej Street, 42-201 Czestochowa, Poland, e-mail: marcinzawa- da04@gmail.com. Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 52 Introduction The harmful effects of climate change made policy makers become in- creasingly interested in reducing greenhouse gas (GHG) emissions using different policy tools such as environmental taxation and regulation impos- ing the increased use of renewable energy. At the international level various steps are taken to motivate countries to reduce the emissions of GHG, e.g. the Kyoto Protocol or the EU energy and climate obligations for member countries. The problem of how the pollutants relate to the economy has been the subject of intense research in the last decades. One of the main develop- ments in understanding the link between the environment and economy was the environmental Kuznets curve (EKC). The term 'environmental Kuznets curve' was coined almost simultaneously by Shafik and Bandyopadhyay (1992), Grossman and Krueger (1995) and Panayotou (1993). It refers by analogy to the inverted U-shaped relationship between the level of economic development and the degree of income inequality formulated by Kuznets (1955). The EKC hypothesis says that environmental degradation increases with per capita income during the early stages of economic growth, and then declines with per capita income after arriving at a threshold1. Hence, the relationship between income per capita and some types of pollutants is ap- proximately an inverted U-shaped. There is a wide stream of researches that has employed cointegration techniques to examine the relationship between per capita income and some types of pollutants, among others, Aspergis and Payne (2009), Halicioglu (2009), Soytas and Sari (2009), Ang (2009), Soytas et al. (2007). However, empirical results are mixed and not conclusive to give policy recommenda- tions that can be applied across countries. The common feature of these stud- ies was the linear approach and symmetric cointegration which may be a possible reason for ambiguous results. It has been suggested more recently (Balke and Fomby, 1997; Enders and Granger, 1998; Enders and Siklos, 2001) that the adjustment of deviations toward the long-run equilibrium need not be symmetric and reverting each period. To our knowledge, there are a very few studies that use non-linear (threshold) cointegration techniques for testing the EKC hypothesis, e.g. Fosten et. al. (2012), Esteve and Tamarit (2012). This paper aims to look at testing the EKC hypothesis for the presence of threshold cointegration between per capita greenhouse gas emissions and per 1 See Stern (2004), Coondoo and Dinda (2002), Dinda (2004), Luzzati and Orisni (2009), Halicioglu (2009) for extensive review surveys of studies which tested the EKC hypothesis. The Environmental Kuznets Curve in Poland – Evidence from Threshold... DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 53 capita income for the Polish economy during the period 2000 to 2012 (quar- terly data). This approach will allow for a different speed of adjustment to the long-run equilibrium depending on whether emissions of greenhouse gas are above or below the EKC. Additionally, another explanatory variable is added, i.e. energy consumption, to test the robustness of the results. The standard EKC hypothesis and its extended version including energy con- sumption will be tested using threshold autoregressive (TAR) and momen- tum threshold autoregressive (MTAR) cointegration method of Enders and Granger (1998) and Enders and Siklos (2001). We will also concentrate on the short-run and long-run causal relationship between per capita greenhouse gas emissions and per capita income using threshold error correction models (T-ECM) and momentum threshold error correction model (M-TECM). To our knowledge, there is no such study that uses this approach for the case of Poland as if the EKC hypothesis is concerned. The remainder of this paper is organized as follows. Section 2 presents the environmental Kuznets curve. Section 3 describes the methodology em- ployed in the analysis. Section 4 describes the data and reports the empirical results. Section 5 concludes. 1. The Environmental Kuznets Curve In classical approach to modelling the relationship between environmen- tal degradation and income level the following quadratic function with the turning point occurring at a maximum pollutant level is used (Agras and Chapman, 1999): ,2210 tttt GDPGDPEP   (1) where tEP – emissions of some pollutant (per capita), tGDP – real income (per capita), both variables are in logarithms, ,1 ,2 0 – estimated param- eters, t error term that may be serially correlated. Based on the parameter values we may conclude about the shape of envi- ronmental pollution and income linkage. If 01  (or 01  ) while 02  , then the function (1) is a monotonically increasing (decreasing) according to linear function behavior. Otherwise, if 01  and 02  , an inverted U-shape describes the situation when the pollution level increases as a coun- try develops until this development reaches a turning point and after that the rising incomes are accompanied by decreasing environmental degradation. The turning point value is approximated by following relation (Stern, 2004): Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 54  .2/exp 21 TPGDP (2) In the EKC literature the more sophisticated functional form is also taken into consideration with the third order polynomial for income factor (Dinda, 2004): .33 2 210 ttttt GDPGDPGDPEP   (3) Similarly to the EKC model (2), the parameter estimates in equation (3) indicate the direction and character of the relationship between environmen- tal pollutant and income. In the case of ,03  ,02  01  and 00  the N-shaped function is monotonically increasing with two possible turning points. In the case of opposite signs of cubic polynomial parameters, namely ,03  ,02  01  and ,00  an inverse-N shape is more accurate for describing analyzed relationship. In order to capture the effect of technological progress on pollution emission level the deterministic time trend (squared time trend) and some additional variables tX that may affect tEP can be included in equation (1) or (3). In our empirical research of the long-run relationship between green- house gas emissions and economic growth first the standard EKC model (1) is assumed and further the EKC model with energy consumption )( tE as additional variable is considered to test the robustness of results. In the latter case the model takes the form: ttttt EGDPGDPEP   2 210 , (4) where  − estimated parameter. 2. Methodology The concept of cointegration implicitly assumes linearity and symmetry, what means that the adjustment of the deviations towards the long-run equi- librium is made instantaneously at each period and increases or decreases of the deviations are corrected in the same way. However, the cointegration tests and their extensions are misspecified if adjustment is asymmetric. To take the property of asymmetry into account, Enders and Siklos (2001) de- veloped the concept of threshold cointegration. This is indeed an extension of residual-based two-stage estimation as developed by Engle and Granger (1987). The differences between them consist in the formulation of linearity and non-linearity from their second stage of unit root test. The Environmental Kuznets Curve in Poland – Evidence from Threshold... DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 55 Extracting from the long-run relationship (1) or (4) the disturbance term t (first stage), in the second stage we focus on the coefficient estimates of 1 and 2 (adjustment parameters) in the following regression:    r i titittttt II 11211 ,)1(  (5) where t is a white noise disturbance. The Heaviside indicator function tI is defined to depend on the lagged val- ues of the residuals t :           1 1 0 1 t t t if if I (6) or on the lagged changes in t :           1 1 0 1 t t t if if I (7) where  is a threshold value. Equations (5)-(6) are referred to as the TAR model (threshold autoregressive model, Enders and Sikolos, 2001), while equations (5) and (7) are named the MTAR model (momentum-threshold autoregressive model; see Enders and Granger, 1998). Petruccelli and Woolford (1984) proved that the necessary and sufficient conditions for the stationarity of residuals  t from the EKC model are ,01  02  and 1)1)(1( 21   for any threshold value  (Enders and Siklos, 2001). If these conditions are satisfied, 0t can be considered as the long run equilibrium value of the sequence. If t is higher than the long-run equilibrium, the adjustment is ,11 t but if t is lower than the long-run equilibrium, the adjustment is .12 t Therefore, the equilibrium error behaves like a threshold autoregressive process (TAR). The MTAR model – according to Enders and Granger (1998) – is especially valuable when adjustment is asymmetric as the deviations t exhibit more 'momen- tum' in one direction than in the other. Hence, the TAR model allows to ex- amine whether the positive deviations )0( t from the long-run equilibri- um have different effects on the behavior of emissions than do the negative deviations ),0( t while the MTAR model allows to display various amounts of autoregressive decay depending on whether the series is increas- Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 56 ing or decreasing (Fosten et al., 2012). There is no prescribed rule as to whether to use the TAR or MTAR model, but it is recommended to select the best adjustment mechanism (TAR or MTAR) using the AIC (Akaike Information Criterion) or SBC (Schwarz Bayesian Criterion) information criteria (Enders, Chumrusphonlert, 2004). In general, the threshold value  governing the asymmetric behavior is unknown and has to be estimated along with the values of adjustment pa- rameters 1 and .2 In our studies we follow Enders and Siklos (2001) and Yau and Nieh (2009) by employing Chan’s (1993) methodology2 of search- ing the consistent estimates of threshold value. However, in many economic applications this value is set to zero, ,0 and then the cointegrating vector coincides with the attractor ( 0t ). Once the threshold value  is obtained and the TAR or MTAR models are estimated, then testing for threshold cointegration can be performed. First, the null hypothesis of no cointegration 0: 210  H is tested, and when it is rejected, then the null hypothesis of symmetric adjustment 210 :  H is verified. To test the null hypothesis of no threshold cointe- gration, Enders and Siklos (2001) proposed the  statistics which is the F statistics. As the distribution of  is non-standard, appropriate critical values were tabulated by Enders and Siklos (2001) and later modified by Wane et al. (2004). In the presence of cointegration (rejection of ),0: 210  H the null hypothesis 210 :  H of symmetric adjustment can be tested using the standard F statistics. When the adjustments coeffi- cients are equal (symmetric adjustment), equation (5) converges the standard ADF test. Rejecting both the null hypotheses of 021   and 21   implies the existence of threshold cointegration with asymmetric adjustment. For instance, if the null of symmetric adjustment is rejected and 21   , then it implies that positive deviations (above threshold) of emissions from the long-run EKC tend to revert quickly towards equilibrium, whereas nega- tive deviations of emissions (below threshold) tend to persist. Hence, it sug- gests faster convergence for positive deviations of emissions (above thresh- 2 The Chan's method to find the consistent estimate of the threshold value arranges the values 1t or 1 t in ascending order, excludes the smallest and largest 15 percent, and the parameter that yields the smallest sum of squared residuals over the remaining 70 percent is the consistent estimate of the threshold. The Environmental Kuznets Curve in Poland – Evidence from Threshold... DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 57 old) than negative deviations (below threshold) from the long-run EKC rela- tionship. Given the threshold cointegration is found, the next step proceeds with the Granger-causality test using the advanced threshold error correction model (TECM) or momentum-threshold error correction model (M-TECM) by Enders and Granger (1998) and Enders and Siklos (2001). The threshold ECM is expressed as the following: ,2 111 1211 321 tit q i iit q i iit q i ittjt vGDPGDPEPZZY              (8) or when additional variable tE is included , 4 321 1 ` 2 111 1211 tit q i i it q i iit q i iit q i ittjt vE GDPGDPEPZZY                   (9) where in equation (8) and (9):  2,, tttjt GDPGDPEPY  , and 11 ˆ     ttt IZ  and ,ˆ)1( 11     ttt IZ  tI – Heaviside indicator function de- termined by (6) or (7), 1ˆ t is obtained from the estimated long-run relation- ship (1) or (4), and tv is a white noise disturbance. Based on the equation (8) and (9) the Granger-causality tests are em- ployed. The long-run causality is determined by the significance of adjust- ment parameters 1 and .2 To test short-run causality (or weak causality) the joint significance of all the coefficients i of ,itEP i of ,itGDP i of 2 itGDP or i of itE  is examined using the Wald F test. It is also desirable to check whether this two sources of causation (short and long-run) are jointly significant. This can be done by providing the Wald F statistics for the interactive terms, i.e. the  terms and the explanatory variables (e.g. ,0: 10  iH ).0: 20  iH The joint test indicates which variables bear the burden of short-run adjustment to re-establish long-run equilibrium, given a shock to the system (Asafu-Adjaye, 2000; Mehrara et al., 2012). This is referred to as the strong Granger causality test. Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 58 3. Empirical Results 3.1. Data Source The data used in this study consist of greenhouse gas emissions ( tEP ) (in tons of CO2 equivalent per capita), real gross domestic product per capita ( tGDP ) and energy consumption 3 ( tE ) in kilo of oil equivalent per capita in Poland. The quarterly data for GDP are obtained from the Central Statistical Office (www.stat.gov.pl) of Poland, while annual data describing green- house gas emissions and energy consumption in Poland were obtained from the Eurostat database. To obtain the real GDP some transformations were made, i.e. quarterly nominal GDP data were transformed by the authors into real GDP in 2005 prices using GDP deflator. Since the GDP data were characterized by signif- icant quarterly seasonality, the TRAMO procedure of Gretl software was applied to adjust it seasonally. The frequency of real GDP series is quarterly, however, the greenhouse gas emissions and energy consumption data are only available at annual frequency. Therefore, we interpolated annual data to quarterly frequency by employing the Denton-Cholette method (Sax, Stei- ner, 2013) in the R software. The sample period is from 2000:Q1 to 2012:Q4. All variables are employed with their natural logarithms form to reduce heteroscedasticity and to obtain the growth rate of the relevant varia- bles by their differenced logarithms. a) annual data b) disaggregated data to quarterly frequency Figure 1. Energy use (in kg of oil equivalent per capita) in Poland in the period 2000–2012 – for annual data and disaggregated data to quarterly frequ- ency 3 Energy consumption covers: consumption by the energy sector itself; distribution and transformation losses; final energy consumption by end users (see epp.eurostat.ec.europa.eu, Eurostat glossary). 2200 2300 2400 2500 2600 2700 2000 2002 2004 2006 2008 2010 2012 k g o f oi l eq u iv a le n t 520 540 560 580 600 620 640 660 680 2 0 0 0 Q 1 2 0 0 1 Q 1 2 0 0 2 Q 1 2 0 0 3 Q 1 2 0 0 4 Q 1 2 0 0 5 Q 1 2 0 0 6 Q 1 2 0 0 7 Q 1 2 0 0 8 Q 1 2 0 0 9 Q 1 2 0 1 0 Q 1 2 0 11 Q 1 2 0 1 2 Q 1 k g o f oi l eq u iv a le n t The Environmental Kuznets Curve in Poland – Evidence from Threshold... DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 59 a) annual data b) disaggregated data to quarterly frequency Figure 2. Greenhouse gas emissions (in tons of CO2 equivalent per capita) in Poland in the period 2000–2012 – for annual data and disaggregated data to quar- terly frequency a) seasonally unadjusted b) seasonally adjusted Figure 3. GDP (in PLN per capita, in 2005 prices) in Poland, 2000:Q1–2012:Q4 The evolution of time series data (original and disaggregated) employed in our analysis is shown in Figure 1–3. 3.2. Cointegration Analysis with Asymmetric Adjustment Before performing cointegration analysis, we use the Augmented Dick- ey-Fuller Generalized Least Squares (ADF-GLS; Elliot et al., 1996) and Kwiatkowski-Phillips-Schmidt-Shinn (KPSS; Kwiatkowski et al., 1992) tests to identify the order of integration for each variable. In Table 1, the ADF-GLS tests show that the unit root hypothesis cannot be rejected at any significant level for each variable in levels. Further investigations of the unit root hypothesis indicate that the first differenced variables are stationary at least at the 10% level of significance. We also apply the KPSS unit root test based on the null hypothesis of stationarity (or no unit root). The results show that the null hypothesis of stationarity is rejected at least at the 10% significance level. Hence, all series are found to be integrated of order I(1). 9,8 10 10,2 10,4 10,6 10,8 11 2000 2002 2004 2006 2008 2010 2012 to n n es o f C O 2 eq u iv a le n t 2,3 2,4 2,5 2,6 2,7 2,8 2 0 0 0 Q 1 2 0 0 1 Q 1 2 0 0 2 Q 1 2 0 0 3 Q 1 2 0 0 4 Q 1 2 0 0 5 Q 1 2 0 0 6 Q 1 2 0 0 7 Q 1 2 0 0 8 Q 1 2 0 0 9 Q 1 2 0 1 0 Q 1 2 0 11 Q 1 2 0 1 2 Q 1 to n n es o f C O 2 eq u iv a le n t 4000 5000 6000 7000 8000 9000 10000 2 0 0 0 Q 1 2 0 0 1 Q 1 2 0 0 2 Q 1 2 0 0 3 Q 1 2 0 0 4 Q 1 2 0 0 5 Q 1 2 0 0 6 Q 1 2 0 0 7 Q 1 2 0 0 8 Q 1 2 0 0 9 Q 1 2 0 1 0 Q 1 2 0 11 Q 1 2 0 1 2 Q 1 P L N 4000 5000 6000 7000 8000 9000 2 0 0 0 Q 1 2 0 0 1 Q 1 2 0 0 2 Q 1 2 0 0 3 Q 1 2 0 0 4 Q 1 2 0 0 5 Q 1 2 0 0 6 Q 1 2 0 0 7 Q 1 2 0 0 8 Q 1 2 0 0 9 Q 1 2 0 1 0 Q 1 2 0 11 Q 1 2 0 1 2 Q 1 P L N Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 60 Table 1. The results of unit root tests Variables Levels Differences ADF-GLS KPSS ADF-GLS KPSS EPt –1.522 [3] 0.463 [2]* –5.246 [3]*** 0.073 [2] GDPt 0.041 [4] 1.218 [4]*** –2.307 [4]** 0.139 [4] GDP2t –0.144 [5] 1.217 [4]*** –1.676 [5]* 0.132 [4] Et –0.957 [2] 1.548 [2]*** –5.368 [2]*** 0.110 [2] Note: (***), (**), (*) in ADF-GLS tests respectively indicate the rejection of the null hypothesis that series has a unit root at 1%, 5% and 10% levels of significance, while in KPSS tests indicate the rejection of the null hypothesis that series is stationary. The numbers inside the brackets are the optimum lag lengths determined using AIC in ADF-GLS tests and the bandwidth is used using the Newey-West meth- od in KPSS tests. Table 2 contains cointegration test results of the standard long-run EKC in the form (1) and Table 3 – long-run EKC including energy consumption in the form (4), when considering threshold (TAR) and momentum adjust- ment (MTAR). The tables report values of the adjustment coefficients 1 and 2 , the  statistics for the null hypothesis of no cointegration (a unit root in t ) against the alternative of cointegration with asymmetric adjust- ment. The F-test is used to test whether the adjustment back to long-run equilibrium is symmetric 21   . It can be seen that all coefficients 21,  have negative signs and are significant at least at 10% significance level. The necessary and sufficient conditions for cointegration hold in the case of all TAR and M-TAR models because the null hypothesis 0: 210  H is rejected (at 5% significance level)4 – see Table 2. Using the standard F-statistics for the restriction 210 :  H it is shown that asymmetric cointegration is strongly significant only in the TAR model with ,0169.0 whereas in the TAR model with 0 the support for asymmetric cointegration is only at 10% significance level. This evidence favors that the adjustment back to equilibrium between greenhouse gas emissions )( tEP and gross domestic product )( tGDP is non- linear. For choosing the more appropriate adjustment process (TAR or M-TAR) we will follow Enders and Chumrusphonlert's (2004) advice in using AIC or SBC to select the best adjustment mechanism. The reported SBC for each model shows that the TAR model with threshold value 0169.0 is more appropriate adjustment mechanism (the minimum SBC is in bold in Table 2). 4 It is worth noting that this null hypothesis is also rejected when critical values from Wane et al. (2004) are taken. The Environmental Kuznets Curve in Poland – Evidence from Threshold... DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 61 Table 2. Results of TAR and M-TAR Enders-Siklos (E-S) test for cointegration on the standard EKC model 1 2  (1=2=0) F (1=2) Lag SBC LB(4) TAR =0 –0.124 (–2.372)** –0.199 (–5.036)*** 12.76*** 3.075 [0.087]* 5 –387.6 5.15 [0.273] M-TAR =0 –0.125 (–2.489)** –0.191 (–4.804)*** 11.722*** 1.722 [0.197] 5 –386.1 6,5 [0.165] TAR =–0.0169 –0.119 (–2.939)** –0.222 (–5.500)*** 15.184*** 6.236 [0.017]** 5 –390.9 4.74 [0.315] M-TAR =0.0023 –0.148 (–2.925)** –0.179 (–4.536)*** 10.722*** 0.416 [0.522] 5 –384.6 5.35 [0.253] Note: *** (**) (*) indicate significance at 1% (5%) (10%) level. Critical values for  statistics from Enders, Siklos (2001). T-statistics for in  parentheses. p-values in brackets. LB(4) for Ljung-Box statis- tics. The lag length is selected such that the AIC is minimized. Table 3. Results of TAR and M-TAR Enders-Siklos (E-S) test for cointegration on the EKC model including energy consumption 1 2  (1=2=0) F (1=2) Lag SBC LB(4) TAR =0 –0.104 (–3.249)** –0.056 (–2.20)* 5.83* 2.155 [0.149] 2 –488.06 6.91 [0.141] M-TAR =0 –0.071 (–2.683)* –0.072 (–2.325)* 4.537 0.002 [0.961] 2 –485.77 7.26 [0.123] TAR =0.0097 –0.115 (–3.516)** –0.055 (–2.237)* 6.616** 3.462* [0.069] 2 –489.40 6.67 [0.155] M-TAR =8.8E-05 –0.064 (–2.36)* –0.082 (–2.74)** 4.74 0.34 [0.563] 2 –486.13 0.03 [0.999] Note: ** (*) indicate significance at 5% (10%) level. Critical values for  statistics from Enders, Siklos (2001). T-statistics for  in parentheses. p-values in brackets. LB(4) for Ljung-Box statistics. The lag length is selected such that the AIC is minimized. We can see in Table 2 that the point estimates 1 and 2 suggest faster convergence for the deviations from long-run EKC equilibrium when they are below the threshold ( 0169.0t ) than when they are above the threshold because 21   . We see that about 22% of the deviation from equilibrium is corrected in the next period when emissions are falling, com- pared to about 12% when they are rising. This means that short-run adjust- ment towards the EKC equilibrium reverts more quickly when the green- house gas emissions are decreasing (below the threshold) and tends to persist more when the greenhouse gas emissions are increasing (above the thresh- old). While the opposite result might be expected, this evidence should not be surprising when we look at the Poland's energy profile. Heavy reliance on Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 62 coal makes Poland a relatively carbon-intensive economy, compared to the IEA Europe average. The large-scale transition of the Polish energy sector, which is characterised by ageing infrastructure5, to a low-carbon economy requires huge long-term investments and an adequate policy and regulatory framework. Therefore, from a short run perspective the potential reduction of greenhouse gas emissions should be rather combined with the energy efficiency improvements. Figure 4. The fitted values of the estimated EKC results for greenhouse gas emis- sions Table 4. Estimated parameters of long-run EKC equation Variable Estimates of parameters (t-statistics) Constant –53.298 (–4.228)*** –23.041 (–2.444)** GDPt 12.225 (4.280)*** 4.345 (1.978)* GDP2t –0.688 (–4.255)*** –0.255 (–2.065)** Et 0.863 (7.639)*** Note: *** (**) (*) indicate significance at 1% (5%) (10%) level. T-statistics for  in parentheses. Estimated parameters for equation (1) and (4). 5 Nearly half of today's operating capacity of the Polish energy sector is older than 30 years. See: Energy Policies of IEA Countries. Poland (2011). The Environmental Kuznets Curve in Poland – Evidence from Threshold... DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 63 Having established the TAR cointegration in the long-run relationship for greenhouse gas emissions, it is now justified to analyse the estimation results (Table 4) and the implications for the EKC hypothesis. We can see that in terms of the coefficients described in the EKC relation (1), we have ,00  01  and ,02  which implies the inverted U-shaped function. The fitted values of greenhouse gas emissions for the observed values of real GDP are displayed in Figure 4. This results show the strong evidence in favour of the EKC hypothesis. The turning point in the observed range of real GDP for greenhouse gas occurs at PLN7218.8 per capita which corre- sponds to around 2007Q2. The addition of energy consumption )( tE to the standard EKC model has not affected the results in terms of the presence of asymmetric cointegra- tion (Table 3), i.e. in the TAR framework (with )0097.0 the asymmetric cointegration is observed but only at 10% significance level when critical values from Enders and Siklos (2001) are used. Moreover, the non-linear relationship remains significant and correctly signed ( 1 and 2 are nega- tive), suggesting the relationship is reasonably robust (see Table 3). Howev- er, when critical values from Wane et al. (2004) are taken, then the null hy- pothesis of non-cointegration )0:( 210  H cannot be rejected. Hence, the results behind the threshold cointegration for the EKC model with energy consumption are rather weak and should be treated with caution. Also the point estimates 1 and 2 differ with regard to the speed of adjustment process and direction of convergence for deviations above and below the threshold value when compared to the standard EKC model (Table 2). Now, faster convergence for deviations (from the long-run EKC) above the threshold than for those below the threshold is observed. We can see (Table 3) that about 12% of the deviation from equilibrium is corrected when emissions are above the threshold ),0097.0( t compared to only 5.5% when they are below the threshold ).0097.0( t Therefore, the short-run adjustment towards the long-run equilibrium reverts more quickly when emissions are increasing (above the threshold) and tends to persist when emissions are decreasing (below the threshold). This result is quite the opposite to that obtained for the standard EKC model what may indicate that energy consumption )( tE is an important determinant of green- house gas emissions. The faster correction for deviations in emissions, if Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 64 they are too high, indicates that in the presence of environmental regulation6 the pressure to reduce them into their long-run levels was occurred. Howev- er, these emissions reductions were to a large extent achieved through the restructuring of Polish industry and energy efficiency improvements. 3.3. Threshold Error Correction Models The positive finding of cointegration with TAR adjustment justifies es- timation of threshold error correction model (8) and (9) and testing the Granger causality7. The T-ECM models are estimated for changes in green- house gas emissions )( tEP and gross domestic product ),( tGDP assum- ing the estimates of threshold values obtained in previous step, i.e. 0169.0 − in model (8) and 0097.0 − in model (9) (see also Ta- ble 2 and 3). The T-ECM for the tE is not estimated since the energy con- sumption is added into the standard EKC model to test the robustness of the results. Based on equations (8) and (9), the Granger causality tests are employed to verify whether all the coefficients of ,itEP itGDP  or 2 itGDP (or itE  in eq. (9)) are jointly statistically different from zero based on a standard F-test (Wald test) and/or whether the coefficients ( 21,  ) of the error correction term are significant. To determine the appropriate lag lengths we apply the SBC criterion, and empirically find that the lag lengths are equal: ,221  qq .332  qq Table 5 presents estimates of the error correction parameters along with Wald F test statistics regarding Granger causality. We will first interpret the results for the T-ECMs and then for T-ECMs including tE (Table 5). While the adjustment speed on the exceeding or underlying threshold level in the T-ECM model for tEP has the 'right' direction ( 1 and 2 have a negative sign and are significant) by acting to eliminate deviations from the long-run equilibrium, the T-ECM model for tGDP adjusts to the 6 Regulation on air pollution has become increasingly stringent, including international protocols such as the Oslo Protocol, the Kyoto Protocol and the EU energy and climate obli- gation for member countries. 7 The T-ECM for 2tGDP is run, but not reported because it has little useful economic in- terpretation (results are available from authors on request). The Environmental Kuznets Curve in Poland – Evidence from Threshold... DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 65 'wrong' direction ( 1 has a positive sign) for one regime, and additionally parameters 1 and 2 are insignificant. In the tEP model the adjustment speed responds faster in the lower regime (emissions below the threshold value) than in the higher regime with increasing deviations from the long-run equilibrium (emissions above the threshold value). This is consistent with the results in Table 2. Now, the greenhouse gas emissions converge to their long-run equilibrium at the rate of 13.2% with a deviation below the thresh- old and at a lower rate of 9.3% with a deviation above the threshold. Table 5. Estimates of threshold error correction model (T-ECM) and the Wald F statistics for Granger causality T-ECM T-ECM including Et EPt GDPt EPt GDPt 1 –0.093* (–2.021) 0.046 (0.693) –0.081 (–1.012) –0.015 (–0.141) 2 –0.132** (–2.610) –0.073 (–1.008) –0.022 (–0.357) –0.163* (–1.961) H0: i=0 1.328 [0.280] 1.519 [0.227] H0: i==0 2.028 [0.111] 1.308 [0.286] H0: i==0 2.952** [0.032] 1.209 [0.324] H0: i=0 1.386 [0.262] 0.562 [0.643] 1.622 [0.202] 1.000 [0.404] H0: i==0 2.059 [0.106] 0.489 [0.744] 1.387 [0.259] 0.750 [0.565] H0: i==0 3.039** [0.029] 0.786 [0.542] 1.293 [0.292] 1.894 [0.133] H0: i=0 5.626** [0.007] 1.162 [0.325] H0: i==0 3.796** [0.018] 1.056 [0.380] H0: i==0 4.487** [0.009] 3.367** [0.029] H0: i=0 3.857** [0.031] 0.129 [0.879] H0: i ==0 2.674* [0.062] 0.187 [0.904] H0: i ==0 2.572* [0.069] 2.012 [0.130] LB(3) 2.60 [0.457] 1.60 [0.659] 2.66 [0.448] 1.95 [0.583] BIC –395.72 –362.00 –387.32 –360.22 Note: *** (**) (*) indicate significance at 1% (5%) (10%) level. t-statistics for 1, 2 in parentheses. p-values for Wald statistics in brackets. LB(3) for Ljung-Box statistics along with p-values. The results of the Granger causality tests show that there is no short-run causality running from real tGDP (and square of tGDP ) to ,tEP but there is an unidirectional short-run causality from greenhouse gas emissions to per capita real .tGDP This can be interpreted as the non-rejection of 0:0 iH  and 0:0 iH  in the tEP equation and rejection of 0:0 iH  in the tGDP equation (see Table 5, 2nd and 3rd column). Besides, the strong Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 66 long-run causality running from per capita real GDP (and square of real GDP) to greenhouse gas emissions is found in the regime below the thresh- old value of −0.00169 (rejection of ,0: 20  iH 0: 20   iH ). The results for the tGDP equation suggest reverse long-run causality running from greenhouse gas emissions to per capita real GDP (rejection of 0: 10  iH and 0: 20  iH ). This implies that deviations from the long-run EKC are corrected not only by movements in greenhouse gas emis- sions but also by movements in per capita real GDP. In the T-ECM model for tEP , including energy consumption (Table 5, 4th column), the parameters 21,  are correctly signed, but statistically in- significant8. Nonetheless, it is worth noting that the correction back to equi- librium is faster in regime above the threshold value of 0.0097 (since )21   unlike the results obtained in the T-ECM not including energy consumption where faster convergence to equilibrium occurred in the regime below the threshold. This means that the addition of energy consumption into the T-ECM changed the response of greenhouse gas emissions to error correction. The manifested influence of energy consumption to reduce greenhouse gas emissions should be rather attached to the promotion of technological progress and the energy efficiency improvements, but not the large-scale transition to low-carbon economy, as has already been men- tioned. Further, there is neither short-run Granger causality (or weak causality) nor long-run causality from tGDP to tEP in the T-ECM including energy consumption (the non-rejection of null hypotheses ,0i ,01  i ,02  i ,0i ,01   i 02   i in Table 5, 4th column). The non-significance of the F-statistics for GDP indicates that it is exogenous in the system. However, there exists the short-run and long-run causality from energy consumption to greenhouse gas emissions (the rejection of null hy- potheses: ,0i 01  i and 02  i in the tEP model including ).tE This evidence suggests that energy consumption has an effect on greenhouse gas emissions and bears the burden of short-run adjustment to restore long-run equilibrium after a shock to the system. In the T-ECM for 8 This finding should be rather attached to the effect of small sample and reduction of de- grees of freedom in the estimated threshold error correction models due to bigger number of variables and lagged terms in comparison to previous TAR estimations (see Table 2 and 3). Therefore the results in Table 5 should be interpreted with caution and validated on the larger sample. The Environmental Kuznets Curve in Poland – Evidence from Threshold... DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 67 tGDP (including energy consumption) the rejection of 0: 20  iH suggests the unidirectional long-run causality from greenhouse gas emis- sions to real GDP in regime below the threshold value of 0.0097. The results from causality analysis based on the T-ECM (including energy consumption) indicate that re-establishing of the long-run equilibrium is carried out by the interaction of greenhouse gas emissions and energy consumption, but the long run effect of GDP is rather weak. Conclusions The paper has aimed to investigate the EKC hypothesis for the case of the Polish economy during the period 2000-2012. We tested for the presence of threshold cointegration between per capita greenhouse gas emissions and per capita income (real GDP). Moreover, to test the robustness of the results, we considered the standard EKC hypothesis with the addition of per capita energy consumption to the model. To address such an issue, we applied the threshold cointegration model which allows the nonlinear adjustment to the long-run equilibrium. In the case of the standard EKC relationship, the results of threshold cointegration indicate that per capita greenhouse gas emissions and per capi- ta real GDP are cointegrated with an asymmetric adjustment process. Ad- justments towards the long-run equilibrium revert more quickly when emis- sions are below the threshold value and tend to persist more when emissions are above the threshold value. The evidence of long persistence of adjust- ments to the equilibrium in higher regime (emissions above the threshold) may be explained by the specific of Polish energy sector, namely heavy reli- ance on coal. As a consequence, the transition to a low-carbon economy will require huge long-term investments and an adequate policy and regulatory framework. Therefore, from a short run perspective the reduction of emis- sions should be rather combined with the energy efficiency improvements. With regard to the Granger causality tests the results are following. In the lower regime the bidirectional long-run causality between per capita real GDP and per capita greenhouse gas emissions is found. The short-run dy- namics suggests that there is no causal relationship from real GDP to green- house gas emissions but there is an unidirectional Granger causality from greenhouse gas emissions to real GDP. Our results find strong evidence in favour of the EKC hypothesis with per capita greenhouse gas emissions having an inverse U-relation with real GDP per capita. The evidence suggests that the turning point in the observed range of real GDP for greenhouse gas emissions occurred at PLN7218.8. As Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 68 a consequence, a decoupling between the two variable appears, i.e. the growth of emissions of some pollutant is slower than the economic growth. The addition of energy consumption to the standard EKC model has not affected the results in terms of the presence of asymmetric cointegration but has changed the direction of convergence for deviations above and below the threshold value. Namely, adjustment towards long-run equilibrium is faster in the higher regime (emissions above the threshold) than in the lower re- gime (emissions below the threshold). This mean that in the presence of environmental regulation the pressure to reduce emissions into their long-run levels is observed. However, it should be emphasized that the abatement of emissions is rather achieved through the restructuring of Polish industry, promoting technological progress and energy efficiency improvements, but not the large-scale transition to low-carbon economy. The results of Granger causality tests indicate that there is no short-run causality from real GDP to greenhouse gas emissions. However, there is short and long-run causality from energy consumption to greenhouse gas emissions. Besides, the unidi- rectional long-run causality from greenhouse gas emissions to real GDP in regime below the threshold value is observed. The results from causality analysis indicate that re-establishing of the long-run EKC is carried out by the interaction of greenhouse gas emissions and energy consumption, but the long run effect of GDP is rather weak. References Acaravci, A., Ozturk, I. (2010), On the Relationship Between Energy Consumption, CO2 Emissions and Economic Growth in Europe, Energy, 35, 5412–5420, DOI: http://dx.doi.org/10.1016/j.energy.2010.07.009. Agras, J., Chapman, D. (1999), A Dynamic Approach to the Environmental Kuznets Curve Hypothesis, Ecological Economics, 28, 267–277, DOI: http://dx.doi.org/10.1016/S0921-8009(98)00040-8. Ang, J. (2007), CO2 Emissions, Energy Consumption, and Output in France, Energy Policy, 35, 4772–78, DOI: http://dx.doi.org/10.1016/j.enpol.2007.03.032. Asafu-Adjaye, J. (2000), The Relationship Between Energy Consumption, Energy Prices and Economic Growth: Time Series Evidence from Asian Developing Countries, Energy Economics, 22, 615–625, DOI: http://dx.doi.org/10.1016/S0140-9883(00)00050-5. Aspergis, N., Payne, J. E. (2009), CO2 Emissions, Energy Usage and Output in Central America, Energy Policy, 37, 3282–86, DOI: http://dx.doi.org/10.1016/j.enpol.2009.03.048. Balke, N. S., Fomby, T. B. (1997), Threshold Cointegration, International Economic Review, 38, 627–645, DOI: http://dx.doi.org/10.2307/2527284. Chan, S.-L. (1993), Consistency and Limiting Distribution of the Least Squares Estimator of a Threshold Autoregressive Model, The Annals of Statistics, 21, 520–533. The Environmental Kuznets Curve in Poland – Evidence from Threshold... DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 69 Coondoo, D, Dinda, S. (2002), Causality between Income and Emissions: a Country Group- Specific Econometric Analysis, Ecological Economics, 40, 351–367, DOI: http://dx.doi.org/10.1016/S0921-8009(01)00280-4. Dinda, D. (2004), Environmental Kuznets Curve Hypothesis: a Survey, Ecological Econom- ics, 49, 431–455, DOI: http://dx.doi.org/10.1016/j.ecolecon.2004.02.011. Elliott, G., Rothenberg, T. J., and Stock, J. H. (1996), Efficient Tests for an Autoregressive Unit Root, Econometrica, 64(4) 813–836, DOI: http://dx.doi.org/10.2307/2171846. Enders, W., Chumrunsphonlert, K. (2004), Threshold Cointegration and Purchasing Power Parity in the Pacific Nations, Applied Economics, 36, 889–896, DOI: http://dx.doi.org/10.1080/0003684042000233104. Engle, F. R., Granger, C. W. J. (1987), Co-integration and Error Correction: Representation, Estimation, and Testing, Econometrica, 55(2), 251–276, DOI: http://dx.doi.org/10.2307/1913236. Enders, W., Granger, C. W. J. (1998), Unit-Root Tests and Asymmetric Adjustment with an Example Using the Term Structure of Interest Rates, Journal of Business Economics & Statistics, 16, 304–311, DOI: http://dx.doi.org/10.2307/1392506. Enders, W., Siklos, P.L. (2001), Cointegration and Threshold Adjustment, Journal of Busi- ness and Economic Statistics, 19, 166–176, DOI: http://dx.doi.org/10.1198/073500101316970395. Energy Policies of IEA Countries. Poland (2011), IEA Review, http://www.iea.org/publications/freepublications/publication/poland2011_web.pdf. Esteve, V., Tamarit, C. (2012), Threshold Cointegration and Nonlinear Adjustment Between CO2 and Income: The Environmental Kuznets Curve in Spain, 1857-2007, Energy Economics, 34, 2148–2156, DOI: http://10.1016/j.eneco.2012.03.001. Fosten, J., Morley, B., Taylor, T. (2012), Dynamic Misspecification in the Environmental Kuznets Curve: Evidence from CO2 and SO2 Emissions in the United Kingdom, Eco- logical Economics, 76, 25–33, DOI: http://dx.doi.org/10.1016/j.ecolecon.2012.01.023. Grossman, G. M., Krueger, A. B. (1995), Economic Growth and the Environment, Quarterly Journal of Economics, 110, 353–377, DOI: http://dx.doi.org/10.2307/2118443. Halicioglu, F. (2009), An Econometric Study of CO2 Emissions, Energy consumption, income and Foreign Trade in Turkey, Energy Policy, 37, 1156–1164, DOI: http://dx.doi.org/10.1016/j.enpol.2008.11.012. Kuznets, S. (1955), Economic Growth and Income Inequality, American Economic Review, 45(1), 1–28. Kwiatkowski, D., Phillips, P.C.B., Schmidt, P., and Shin, Y. (1992), Testing the Null Hypoth- esis of Stationarity Against the Alternative of a Unit Root, Journal of Econometrics, 54, 159–178, DOI: http://dx.doi.org/10.1016/0304-4076(92)90104-Y. Luzzati, T., Orsini, M. (2009), Investigating the Energy-Environmental Kuznets Curve, Ener- gy, 34, 291–300, DOI: http://dx.doi.org/10.1016/j.energy.2008.07.006. Mehrara, M., Musai, M., Nasibparast, S. (2012), The Causality between Savings and GDP in Iran, International Journal of Advanced Research in Engineering and Applied Scienc- es, 1(6), 43–55. Panayotou, T. (1993), Empirical Tests and Policy Analysis of Environmental Degradation at Different Stages of Economic Development, World Employment Programme Working Paper Number WP238, International Labor Office, Geneva. Petruccelli, J. Woolford, S. W. (1984), A Threshold AR(1) Model, Journal of Applied Proba- bility, 21(2), 270–286, DOI: http://dx.doi.org/10.2307/3213639. Sax, C., Steiner, P. (2013), Temporal Disaggregation of Time Series, The R Journal, 5(2), 80– –88, http://journal.r-project.org/archive/2013-2/sax-steiner.pdf. Mariola Piłatowska, Aneta Włodarczyk and Marcin Zawada DYNAMIC ECONOMETRIC MODELS 14 (2014) 51–70 70 Shafik, N., Bandyopadhyay, S. (1992), Economic Growth and Environmental Quality: Time Series and Cross-Country Evidence, Background Paper for the World Development Report 1992, World Bank, Washington D.C. Soytas, U., Sari, R. (2009), Energy Consumption, Economic Growth and Carbon Emission: challenges Faced by an EU Candidate Member, Ecological Economics, 68, 1667–75, DOI: http://dx.doi.org/10.1016/j.ecolecon.2007.06.014. Soytas, U., Sari, R., Ewing, B.T. (2007), Energy Consumption, Income and Carbon Emissions in the United Stated, Ecological Economics, 62, 482–489, DOI: http://dx.doi.org/10.1016/j.ecolecon.2006.07.009. Stern, D. (2004), The Rise and Fall of the Environmental Kuznets curve, World Development, 32, 1419–1439, DOI: http://dx.doi.org/10.1016/j.worlddev.2004.03.004. Wane, K. M., Gilbert, S., Dibooglu, S. (2004), Critical Values of the Empirical F-Distribution for Threshold Autoregressive and Momentum Threshold Autoregressive Models, 2004 Discussion Papers for the Department of Economics, Southern Illinois University at Carbondale. Yau, H.-Y., Nieh, C.-C. (2009), Testing for cointegration with threshold effect between stock prices and exchange rates in Japan and Taiwan, Japan and World Economy, 21, 292– –300, DOI: http://dx.doi.org/10.1016/j.japwor.2008.09.001. Ekologiczna krzywa Kuznetsa dla Polski − analiza progowej kointegracji Z a r y s t r e ś c i. Artykuł przedstawia analizę relacji długookresowej między emisją gazów cieplarnianych w przeliczeniu na mieszkańca a realnym PKB w przeliczeniu na mieszkańca (hipoteza EKC, ekologiczna krzywa Kuznetsa) z wykorzystaniem podejścia progowej (asy- metrycznej) kointegracji i modelu korekty błędem dla polskiej gospodarki w okresie 2000- 2012 (dane kwartalne). Standardowy model EKC został rozszerzony o zużycie energii w celu zbadania wpływu dodatkowych zmiennych na wyniki. Hipoteza EKC była sprawdzana z wykorzystaniem progowych modeli autoregresyjnych (TAR i MTAR). Ponadto, dla zbada- nia krótkookresowej i długookresowej przyczynowości Grangera między emisją gazów cie- plarnianych w przeliczeniu na mieszkańca a realnym PKB w przeliczeniu na mieszkańca zastosowano progowy model korekty błędem. Otrzymane wyniki dostarczają istotnych dowo- dów na rzecz hipotezy EKC dla przypadku Polski oraz wskazują, że tymczasowe odchylenia od ścieżki długookresowej równowagi EKC są korygowane w asymetryczny sposób. S ł o w a k l u c z o w e: ekologiczna krzywa Kuznetsa, emisja gazów cieplarnianych, zużycie energii, wzrost gospodarczy, kointegracja progowa, przyczynowość w sensie Grangera.