Environmental efficiency evaluation of China s industrial system. considering the learning effects

Transcription

1 Environmental efficiency evaluation of China s industrial system considering the learning effects By Kangjuan Lv Yiwen Bian Anyu Yu* SHU-UTS SILC (Sydney Institute of Language & Commerce) Business School Shanghai University, Shanghai, , P.R. China. School of Economics and Management Tongji University, Shanghai , P.R. China. yuanyu1990yy@163.com * Corresponding author: Mr. Anyu Yu School of Economics and Management Tongji University Shanghai , P.R. China Phone: (86) Fax: (86) address: yuanyu1990yy@163.com 1

2 Astract The learning effect is an important source of organizational performance improvement, which is driven y the cumulative learning over time. However, in the literature, the learning effect on the environmental efficiency improvement is ignored. We in this paper aim to examine the environmental efficiency of industrial system in China y considering the learning effects on the reductions of waste gas emissions. To this end, an improved SBM model is proposed y incorporating the learning effects. These learning effects are estimated ased on the log-linear learning curve model. The proposed approach is applied to evaluate the environmental efficiency of China regional industrial production systems in The application results show that the learning effects have significant impacts on the waste gas emissions as well as the environmental efficiency. We also find that the proposed approach accompanied with the traditional DEA model can effectively help us to identify the inefficiency whether sourced from the learning activities, other production activities, or the oth. It is interestingly found that, some regions should take more efforts to improve the learning performance, while some others should enhance their production activities so as to improve their environmental efficiency. In particular, some regions should e cautious in spending more resources on the learning activities. Some other findings and insights are also achieved. Keywords: Environmental efficiency; Learning effect; Slacks-ased measure approach; Industrial system; China 2

3 1. Introduction Nowadays, China has een regarded as the gloal factory, which is due to the great achievements of China's industrial system during the past 30 years (Choi and Zhang, 2011). However, this achievement is seriously at the cost of environmental quality. For example, waste gas emissions from China s industrial system have increased from illion m 3 in 2000 to illion m 3 in 2014 (National Bureau of statistics of China, ). As such, environmental and ecological protection has attracted increasing concerns in recent years in China. In order to reduce industrial waste gas emissions, Chinese government has in recent years implemented various policies and strategies, e.g., restricting the expansions of energy-intensive companies, increasing the proportion of non-fossil fuels in the energy consumption to 11.4%, and technology upgrading for desulfurization facilities. It is well known that, economic growth in China is mainly driven y energy consumption, and energy consumption in China is significantly dominated y fossil fuels, i.e., coal, crude oil and natural gas, which are the main causes of waste gas emissions in the country as well as in its industrial systems. In such a circumstance, it is etter for the government to improve environmental efficiency y simultaneously considering waste gas emissions and energy consumption as well as other production factors such as laor, capital and GDP (gross domestic product). Therefore, environmental efficiency improvement might e a cost-effective way to effectively reduce waste gas emissions, increase energy security and keep economic growth. Learning can e defined as an organization s capaility to improve its performance ased on the past experience as well as knowledge and understanding (Chang et al., 2013; Kuo and Yang, 2006; Garvin, 1993). As Chang et al. (2013) summarized, the learning process involves the acquisition of various knowledge, knowledge sharing and use, and an organization can create learning opportunities through internal knowledge transfer and new knowledge generation. Learning opportunities can provide effective environments for an organization s memers to innovate (Tsai, 2001), and continuous learning might enale an organization to adjust its production and management activities y trade-offing organizational inputs, outputs and the environment (Senge, 1990). Many studies have also suggested that the performance can e effectively improved y continuous organizational learning over time (Sorenson, 2003; Chang et al., 2013). 3

4 In the environmental management practice, China has taken much efforts to provide opportunities for various organizations, local governments or firms to create learning environments, e.g., sending various organizations managers and core professionals and technicians to study aroad. At the same time, China has invested much to further promote the reductions of industrial waste. During 2000 to 2014, the special investment to industrial waste gas emission aatements has sharply increased from illion RMB Yuan to illion RMB Yuan (National Bureau of statistics of China, 2015). As Calantone et al. (2002) indicated, this kind of input can encourage innovation. This might effectively help organizational memers to optimize their ehaviors to achieve the optimal performance (Garcia-Morales et al., 2007). Thus, more learning input might lead to etter organizational performance (Hsu and Pereira, 2008). From these studies and evidences, we can conclude that, much investments to the waste gas emission reductions can effectively improve the environmental efficiency of China s industrial system. Environmental efficiency has een regarded as an effective way to measure the aggregate economic and environmental performance with the consideration of energy consumption, which can offer condensed information for managers to cope with environmental issues (Zhou et al., 2008a; Bian et al., 2015). In the literature, various approaches have een proposed to measure environmental efficiency. A common practice is to first select different suitale and practical indicators and then integrate them into a composite efficiency index. The extant efficiency evaluation methods are generally grouped into parametric and non-parametric methods (Sadjadi and Omrani, 2008). Parametric methods such as stochastic frontier analysis evaluate performance through estimating a cost or production function. Therefore, deviations in function forms would affect the results of such methods (Sadjadi and Omrani, 2008). By contrast, it is not necessary to estimate cost or production functions when using non-parametric approaches such as data envelopment analysis (DEA), and thus these approaches can effectively avoid model misspecification (Wei et al., 2007; Choi et al., 2012). For these reasons, DEA in recent years has een widely used to measure regional or national environmental efficiency (Zhou et al., 2008a). The existing DEA approaches for environmental efficiency evaluation can e generally grouped into five categories. The first is a hyperolic measure approach proposed y Färe et al. (1989), which is a nonlinear DEA model y using a reciprocal (hyperolic) measure to evaluate 4

5 the efficiency with undesirale outputs. The second treats undesirale outputs (pollutants) as inputs in traditional DEA models (Hailu and Veeman, 2001; Hu and Wang, 2006). The third uses a simple data transformation function to translate undesirale outputs into normal outputs. For instances, Seiford and Zhu (2002) changed the undesirale outputs to e positive desirale outputs y using a linear monotone decreasing transformation. The fourth is a directional distance function approach ased on the concept of weak disposaility technology, which evaluates and improves the DMU s efficiency according to the given efficiency improvement direction (Färe and Grosskopf, 2004). The last one is a slacks-ased measure (SBM) DEA approach proposed y Tone (2001), and applied y Zhou et al. (2006) to estimate environmental efficiency. Choi et al. (2012) used this approach to examine the efficiency and aatement costs of energy-related CO 2 emissions in China. This approach evaluates the efficiency in consideration of slacks values of all inputs and outputs. Based on these approaches, many studies have further examined the issues of caron emissions performance (Lu et al., 2013), pollutants reduction cost estimation (Wang et al., 2015) and environmental efficiency with pollutants or emissions uncertainties (Jin et al., 2014; Zha et al., 2016) as well as energy efficiency in consideration of pollutants (Wang et al., 2013). A growing numer of studies in the literature have shown that the impact of learning on organizational performance improvement can e predicted and examined using simulation learning, which has een widely used in usiness, science, technology, engineering and industrial context (Badiru and Ijaduola, 2009). These studies have suggested that, regression approach is the most widely used method in investigating the impact of learning effect on organizational performance (Pramongkit et al., 2000; Badiru and Ijaduola, 2009; Chen and Chang, 2010; Grosse et al., 2015). To the est of our knowledge, the extant studies on the effects of learning on organizational performance under the DEA framework are limited. Chang et al. (2013) has incorporated learning effect into DEA models to measure the performance of municipal solid waste recycling system in Taiwan. In their work, the learning effect on municipal solid waste recycling performance is estimated y using a regression approach. Azadeh et al. (2013) integrates fuzzy data envelopment analysis and fuzzy simulation approach to optimize the operator allocation in multiproduct cellular manufacturing systems y considering learning effect. They estimate the learning effect y using processing times y operators. Note that, these two studies mainly focus on examining the impacts of learning effects on desirale outputs or desirale inputs 5

6 rather than undesirale outputs such as waste gas emissions. Furthermore, these studies have incorporated the learning effects into DEA models y treating them as an independent indicator and thus a constraint. However, in real production operational practice, the learning effects can also e represented as some practical and meaningful factors such as reduced or increased cost, pollutants or energies. We can conclude from the aove mentioned studies that, only limited studies have examined the impacts of learning effects on organizational performance under DEA framework y considering desirale outputs rather than undesirale outputs. Also, some studies have investigated the influences of learning effects only on caron reductions, e.g., Lohwasser and Madlener (2012) and Yu et al. (2015). Environmental efficiency as an overall performance measure is not explored in these studies. The primary goal of this paper is to investigate the impact of learning effect on the environmental efficiency of regional industrial systems in China. In practice, the impact of learning effect on the industrial system environmental efficiency can e illustrated as the following three aspects. First, continuous learning for acquiring more knowledge and new knowledge can assist to adjust industrial system s operations to achieve the equilirium etween inputs, desirale outputs such as GDP as well as the pollutants. Second, learning can effectively encourage the innovations in production technologies, cleaner production technologies, production managements as well as the corresponding applications to the practice. Third, managers, administrative staffs and engineers can acquire experiences through industrial production and management practices, which can result in more energy saving and less pollutants discharges. These impacts raise the following three academic research issues: (1) How to estimate the impacts of learning effects on pollutants reductions? (2) How to evaluate environmental efficiency y considering the learning effects? (3) How to measure potential pollutants reductions? These issues should e effectively addressed efore taking some actions to reduce the waste discharges in China s industrial systems. In order to effectively measure environmental efficiency of industrial systems in China with the consideration of learning effects, we in the current paper propose an improved slacks-ased DEA approach. In the descried approach, learning effects are represented as the reductions of the undesirale outputs, which are estimated y using a regression approach. We then incorporate these effects into the corresponding undesirale outputs. The proposed approach is used to 6

7 evaluate environmental efficiency of China s regional industrial systems in The remainder of the paper is organized as follows. Section 2 firstly introduces the asic SBM model for measuring environmental efficiency, and the learning effect estimation approach. Then, the proposed model is provided. The measures of potential industrial waste reductions y considering learning effects are also defined. In Section 3, the proposed approach is applied to examine environmental efficiency of regional industrial systems in China with 2010 dataset. Conclusions and some policies suggestions are offered in Section Methodology Consider that there are n independent regional industrial systems in China, denoted y R j ( j 1,2,...,n ). In the process of industrial production, each regional industrial system employs m inputs x ( i 1,2,..., m) such as laor, capital and energy as inputs to produce k ij desirale outputs y ( r 1,2,..., k ) such as GDP along with p generated undesirale g rj outputs y ( l 1,2,..., p) such as industrial sulfur dioxide emission (SO 2 ), industrial soot emission and industrial dust emission. 2.1 The asic SBM model In the literature, various traditional DEA or directional distance function approaches have een proposed to evaluate energy efficiency and environmental efficiency in the literature (Zhou et al. 2008). However, the efficiency measures in these models are all radial measures, which allow for the equal and proportionate adjustments for all inputs or outputs. Thus, these approaches cannot identify the inefficiency information for specific inputs or outputs, and this in turn may result in a iased estimation (Fukuyama and Weer 2009). To overcome the weakness of radial measure, Zhou et al. (2006) developed a SBM model to measure the environmental performance, and Choi et al. (2012) applied the SBM model to measure the environmental efficiency and aatement costs of energy-related CO 2 emissions in China. Following these work, we modify the standard SBM model (Tone, 2001) y incorporating the slacks of undesirale outputs as 7

8 0 m 1 si 1 m x min 1 1 ( ) k p y y n j 1 n j 1 n j 1 j ij i i0 g g j rj r r 0 j l l 0 i 1 i0 k p sr sl g r 1 r0 l 1 l0 s.t. x s x, i 1,..., m, y s y, r 1,..., k, y s y, l 1,..., p,, s, s, s 0, j 1,2,..., n. j i r l (1) Note that, s i, s r and s l are slacks attached to inputs, desirale outputs, and undesirale outputs, respectively. The suscript 0 represents the regional industrial system to e evaluated. j represents an intensity variale that indicates to what extent the particular R j is involved in the production. It is noteworthy that, model (1) is a CCR model with constant returns to scale (CRS), which can capture the overall technical efficiency (pure technical efficiency and scale efficiency) of the evaluated regional industrial system. According to Zhou and Ang (2008), the CRS assumption can satisfy all production technologies. Thus, we in this paper use the CCR model to examine regional industrial systems efficiencies. Note that, the government have invested or attracted special capital to reduce industrial waste in practice. Thus, for ease of analysis, we in this paper assume that the aatements of undesirale outputs are conducted separately rather than jointly with the production process. Thus, undesirale outputs in our study are assumed to e freely or strongly disposale. Since model (1) is a non-linear programming, following Tone (2001), it can e transformed into the following linear model: 8

9 0 m 1 Si min( q ) m x 1 S S s.t. q ( ) 1 k p y y n j 1 n j 1 n j 1 i 1 i0 k p r l g r 1 r0 l 1 l0 x S qx, i 1,..., m, j i j i i0 y S qy, r 1,..., k, g g j rj r r 0 y S qy, l 1,..., p, j l l 0, S, S, S 0, j 1,2,..., n. j i r l (2) In model (2), k q k, Si qs i, Sr qs r and Sl qs l are the transformed slack variales corresponding to k, s i, s r and s l in model (1), respectively. 0 is the overall efficiency of current under evaluated regional industrial system. By solving model (2), the optimal values of all slacks and q can also e otained. If all the slacks are equal to zero, the evaluated regional industrial system is efficient in the overall efficiency; otherwise, it is inefficient. Note that, model (2) can measure a regional industrial system s overall efficiency y considering the common inputs, desirale outputs and undesirale outputs. However, learning effects are not addressed in this model. We in what follows firstly present an approach to estimate the impacts of learning effects on the reductions in undesirale outputs, and then incorporate the learning effects into the asic SBM model. 2.2 Estimation of learning effects Wright (1936) is the pioneer on examining the impact of learning effect on the cost of airplanes and modeling it as a log-linear function of the cumulative unit numer. In his work, the proposed model (also called the log-linear model) uses a log-linear function to characterize the learning curve. Based on his work, various models have een proposed to estimate learning effects, e.g., S-curve model, exponential model and hyperolic model (Grosse et al., 2015). Among the existing models, the log-linear model has een widely used and also een regarded as the asic model to measure the cumulative learning effects (Badiru and Ijaduola, 2009; Chang et al., 2013). 9

10 This might e due to its simplicity and generally good fit to the oserved data (Nemet and Husmann, 2012). Most of the existing studies use regression approaches to explain the influence degrees of the learning effects ased on this model (Badiru and Ijaduola, 2009). In recent years, learning curve has een frequently adopted to address the environmental or energy issues. Specifically, ased on traditional learning curves, Yu et al. (2015) has adjusted the Co-Douglas multiplicative exponential model to descrie caron intensity learning curve, which is used to characterize the trend of caron intensity with respect to per capita GDP changes. They also apply the technological learning curve to model energy-technology diffusion and policies. Similar study can also e found in Lohwasser and Madlener (2012). In a recent work, Chang et al. (2013) has proposed a DEA model y considering the cumulative learning effect to examine municipal solid waste recycling performance. We in this paper also following the prior work apply the regression models to estimate the impacts of learning effects on the reductions of undesirale outputs. To this end, we firstly extend the primary learning curve model proposed y Wright (1936) to the following model: t EI EI ( z ) ( d 0, l 1,2,..., p, j 1,2,..., n ). (3) t 0 s d s 1 Note that in Eq.(3), t EI is the pollutant intensity of undesirale output y for regional industrial production system j at time t, and 0 EI represents the initial pollutant intensity in the industrial production process at the eginning of the study time period or efore the time period 1. s z is the independent variale at time s, which represents the pollutant reduction volume corresponding to time s; and thus t s 1 z refers to the cumulative volume of pollutant s reduction during the whole study time period. d in this equation is the learning coefficient. Eq. (3) suggests that regional industrial system s pollutants intensities can e reduced y the increase of cumulative industrial pollutants reductions. This pollutant reduction process can e realized y experience accumulation and technology advancement (Yu et al., 2015). Thus, Eq. (3) also indicates that the learning effects can help to reduce industrial pollutants intensities in regional industrial production process. It is noteworthy that, Wright (1936) in this work aims to examine the impact of learning 10

11 effect on the production cost. Thus, he uses the cumulative production volume during his study time period to estimate the production cost. It is stated that, we in this paper use the cumulative pollutant reduction amount to investigate the learning effect on each pollutant intensity in industrial production system. In fact, the pollutant intensity is defined as the ratio of pollutant amount to GDP value, which can also e seen as a kind of production cost. Similar assumption can also e found in Yu et al. (2015). Note also that, following Wright (1936), we argue that the learning effect at time t also affects the pollutant intensity at this time, and thus we use the cumulative pollutant reduction amount during the period from time 1 to time t rather than from time 1 to time t-1, which is different from that in Chang et al. (2013). Similar treatments of the time periods can also e found in Ferioli et al. (2009) and Yu et al. (2015). In order to measure the impact of learning effects on pollutants intensities, following the existing studies such as Badiru and Ijaduola (2009) and Chang et al. (2013), we take the natural logarithm on oth sides of Eq.(3), and have the following learning curve model: t t 0 s ln s 1 ln EI EI - dln z. (4) Regarding the impact of organizational learning on performance improvement, most studies have focused on the endogenous variales of the learning coefficient (d) effects on the performance (Chang et al., 2013). The endogenous variales may remain constant over time during a particular time period for the logged ratio of production at the end of the time period relative to that at the eginning of the period. This might also e due to the fact that, when the jo or task related skills are routine and easy to learn, organization s memers do not easily forget the cumulative experiences (Yang and Chang, 2008). Thus, we following Chang et al. (2013) assume that the pollutant reduction technologies and experiences could not e forgotten, and the learning effects of regional industrial production keep constant during the study time period. In this case, the learning effects on pollutant intensity are only influenced y the endogenous variales s z. As such, ased on the prior work such as Chang et al. (2013), the learning coefficient of d can e estimated y using a regression approach. According to the aove mentioned statements, the cumulative pollutants reductions caused y the learning effects during the study time period can e otained y using the following 11

12 equation: EIR EI EI. (5) t 0 t Note that, t EIR refers to the cumulative pollutant reduction of y for regional industrial system j at time t caused y the learning effect during the study time period. Eq. (5) can e expressed as the following equation: t EIR EI EI ( z ). (6) t 0 0 s d s 1 Note that in Eq.(6), the pollutant intensity reduction t EIR decreases in the cumulative pollutant reduction during the time period. Thus, the possile pollutant reduction at time t can e otained y t t t ER EIR GDP j. (7) Note that in Eq.(7), t GDP j is the GDP output of regional industrial system j at time t. Different from Yu et al. (2015), we in this study estimate the possile reductions for the undesirale outputs rather than inputs y taking the learning effects into consideration. We will calculate the amounts of pollutants at time t for each regional industrial systems y adding the corresponding possile reductions to their real discharged amounts at time t. In this circumstance, t ER can e seen as the reductions caused y the learning effects. For ease of notations, we set t y = ER at time t. These recalculated pollutants amounts are incorporated into the asic DEA model, i.e., model (2), in what follows. 2.3 The proposed SBM model with learning effects Learning effect can affect the cumulative pollutant discharge over time, and thus it can e measured using time series data. Note that, continuous learning can effectively improve an organizational performance over time, and this improvement in our study is represented y the pollutants reductions in regional industrial systems, which is achieved y the continuous learning from accumulating more knowledge, investing more intellectual capitals, promoting memers ehavioral changes, enaling organizations to increase desirale outputs, technology improvements as well as management performance enhancements (Chang et al., 2013). 12

13 We can conclude from these statements that, the real discharged pollutants at time t have een considered the impacts of the learning effects during the study time period. Thus, to differentiate these impacts, the possile discharge amount for each undesirale output at time t can e formulated as y y. By incorporating these indicators into model (2) instead of the original undesirale outputs, we have the following model: 0 m 1 Si min( q ) m x 1 S S s.t. q ( ) 1 k p y y y n j 1 n j 1 n j 1 i 1 i0 k p r l g r 1 r 0 l 1 l 0 l 0 x S qx, i 1,..., m, j i j i i0 y S qy, r 1,..., k, g g j rj r r 0 ( y y ) S q( y y ), l 1,..., p, j l l0 l0, S, S, S 0, j 1,2,..., n. j i r l (8) Note that, model (8) can exclude the impacts of the learning effects from the environmental efficiency evaluation compared with model (2). Chang et al. (2013) in their study incorporated the learning effects into the DEA model y treating the effects as independent inputs, which is different from ours. The potential assumptions in their work are that the learning effects can e optimized in the process of efficiency evaluation. However, this might not e practical, for that we cannot change the learning performance in past years. Thus, we argue that the learning effects cannot e incorporated into the efficiency evaluation models as inputs. As this, we in this study use the recalculated undesirale outputs instead of the original undesirale outputs to estimate the impacts of the learning effects on undesirale outputs as well as the environmental efficiency. The overall efficiency 0 of regional industrial production system can e otained y solving model (8), and the optimal values of * q and S, * i S and * r S can also e acquired * l at the same time. If all the optimal slacks are evaluated as zero, the under evaluated regional industrial system is said to e efficient in the overall efficiency; otherwise, it is inefficient. For an inefficient regional industrial system, its efficiency can e improved y reducing the slacks of inputs and (or) outputs. Therefore, we have the following equation to calculate the 13

14 potential pollutants reductions for the under evaluated regional industrial system R 0 : GP S q * l0 l0 * ( l 1,..., p). (9) Note that, GP l0 refers to the potential reduction of y l 0 ( l 1,..., p ) with the consideration of the learning effects under model (8). Following Hu et al. (2006), the environmental efficiency for each pollutant can e defined as the ratio of its target value to its real value. This approach has een widely used in related studies in recent years, e.g., Wang et al. (2015). Thus, we have the following environmental efficiency measure for each pollutant of R 0 : EE ( y y ) GP l0 l0 l0 l0 yl0 yl0 ( l 1,..., p). (10) We can see in Eq.(10) that, EE l0 is the efficiency index for the pollutant y l 0 of R 0 y considering the learning effects. It is noted that, we can also define the efficiency measures for all pollutants under model (2), which can e used to estimate all pollutants efficiencies without the consideration of the learning effects. For ease of notations, denote reduction potentials in the pollutants y GP l0, and efficiency measures for the pollutants y EE l0 ( l 1,..., p ), respectively. These measures are similar to those defined in Eq.(9) and Eq.(10), respectively, and thus omitted here. In the following section, we will apply the proposed approach to examine the environmental efficiency of regional industrial systems in China in The empirical study 3.1 Regions and the data source There are 31 regions (provinces, autonomous regions and municipalities) in mainland China, which can e divided into three groups, i.e., the eastern, central and the western areas. Similar classifications can also e found in Bian et al. (2015). These regions are depicted in Fig.1. 14

15 Fig.1. Regions and areas in China As shown in Fig.1, the eastern area contains eight coastal provinces such as Shandong, Jiangsu, Zhejiang and Guangdong as well as three municipalities, i.e., Beijing, Tianjin and Shanghai. This area is the most developed area in the country. For example, its industrial GDP accounts for aout 57.48% of China s whole industrial GDP in 2010 (National Bureau of statistics of China, 2011). The central area includes 10 inland provinces such as Heilongjiang, Inner Mongolia, Anhui, Huei and Hunan, which is a home ase for agriculture and related industries. This area also has some heavy industrial ases such as Heilongjiang. The western area consists of nine provinces such as Gansu, Qinghai, Xinjiang and Sichuan as well as one municipality (i.e., Chongqing). This area has the lowest population density and is also the least developed area in China. Its GDP only accounts for aout 12.85% of the total GDP in China in 2010 (National Bureau of statistics of China, 2011). Note that, since the data on energy input and some pollutants of the industrial system of Tiet is not availale, it is excluded and 30 regions are considered in this study. We in this study take laor, capital stock and energy consumption as three inputs, industrial added value as one desirale output, and SO 2 emission, soot emission and dust emission as three undesirale outputs. Due to the unavailaility of some data efore 1997, especially for the data on Chongqing, we take as the study time period. The data on laor and capital stock in 2010 are collected from Statistical Yearook of China in The data on energy consumption 15

16 of regional industrial systems in 2010 is collected from China Energy Statistical Yearook in The data on industrial value added and three undesirale outputs of regional industrial systems is collected from Statistical Yearook of China during It is noteworthy that, to calculate the pollutants intensities, we convert the data of industrial value added into real one at 1997 price. When evaluating the environmental efficiency, the data on industrial value added is its nominal value at The descriptive statistics of the data set in 2010 is summarized in Tale 1. Tale 1. Descriptive statistics Indicators Unit Max Min Mean Std.Dev Inputs Energy 10 4 tons SCE a Laor 10 4 people Capital Billion Yuan Desirale output Industrial added value Billion Yuan Undesirale outputs SO tons Soot 10 4 tons Dust 10 4 tons a Note: SCE refers to standard coal equivalent. 3.2 Estimation of learning effects The prior literature such as Grosse et al. (2015) suggests that the ordinary least squares (OLS) is widely used to estimate the learning effects ased on learning curve models. Following the existing studies, we use OLS regression model to estimate the learning effect coefficients for the three waste gas emissions, i.e., SO 2, soot and dust. Then ased on Eq. (7), the possile reductions of the three waste gas emissions caused y the learning effects can e otained. The results are shown in Tale 2. Note that in Tale 2, SRL, STRL and DRL are used to represent the possile reductions in SO 2, soot and dust, respectively. We can find in Tale 1 and Tale 2 that, mean value of possile SO 2 reduction caused y the learning effects of all regional systems is ton during , which is much larger than that of the real emission in 2010, i.e., ton. That is, the possile total SO 2 emission volume is ton in 2010 y excluding the influence of the learning effects. This indicates that the possile reduction in SO 2 accounts for aout 78.56% of the total possile emission. Similarly, we can otain the proportions with respect to soot and dust, i.e., 86.05% and 86.90%, respectively. These results mean that, on the one hand, all the waste gas 16

17 emissions are largely affected y the learning effects; and on the other hand, the learning effects have higher impacts on dust than those on soot and SO 2. In particular, the learning effects have the lowest impacts on SO 2 emission among the three waste gas emissions. Tale 2. Reductions in the waste gas emissions Regions SRL (10 4 tons ) STRL (10 4 tons ) DRL (10 4 tons ) Eastern Beijing area Tianjin Heei Liaoning Shanghai Jiangsu Zhejiang Fujian Shandong Guangdong Hainan Central Shanxi area Inner M a Jilin Heilongjiang Anhui Jiangxi Henan Huei Hunan Guangxi Western Chongqing area Sichuan Guizhou Yunnan Shaanxi Gansu Qinghai Ningxia Xinjiang Mean a Note: Inner M refers to Inner Mongolia. This will also e used in the remaining part in this study. It is easily found in Tale 2 that, there are remarkale geographic disparities in the learning 17

18 effects on waste gas emissions. The western area is estimated to have the lowest mean possile reductions in the three waste gas emissions, i.e., , and tons for SO 2, soot and dust, respectively. The eastern area has the highest mean possile SO 2 reduction ( ton), and the central area has the highest mean possile soot and dust reductions, i.e., and tons, respectively. These disparities can also e found etween regional industrial systems. The learning effects on the waste gas emissions of some regional industrial systems are much higher than those of others. For example, the possile reduction in SO 2 of Xinjiang is ton, which is much lower than its real emission in 2010 ( ton). However, this is just the opposite in Ningxia, whose real SO 2 in 2010 is ton and the possile reduction during the study time period is ton. All the aove results indicate that, the learning effects have significant impacts on waste gas emissions of national industrial systems, and there are significant unalances etween the learning effects among regional industrial systems. These results also suggest that the learning effects might have significant impacts on the environmental efficiency of the industrial system in China. 3.3 Efficiency analysis We in this su-section firstly compare the overall efficiency results otained from model (2) and model (8) to illustrate the rationality of the proposed approach, and then use the proposed approach to examine the environmental efficiency of regional industrial systems in China. The efficiency results are reported in Tale 3. It can e found in Tale 3 that, there are eight regions, i.e., Beijing, Tianjin, Shanghai, Fujian, Guangdong, Inner Mongolia, Henan and Chongqing, are evaluated as efficient in the overall efficiency of their industrial systems. However, under the proposed approach, nine regions are rated as efficient. In particular, Beijing shifts from efficient under model (2) to inefficient under model (8), while Jilin and Heilongjiang shift from inefficient under model (2) to efficient under model (8). The efficiency differences etween these two models can also e oserved in mean overall efficiencies under the two models. Note that, mean overall efficiency of regional industrial systems under model (8) is , which is larger than that under model (2), i.e., These differences are caused y the consideration of the learning effects in the efficiency evaluation. It is noted that, mean overall efficiency scores under the two models are oth not very high. This 18

19 means that, the overall environmental efficiency of national industrial system is relatively low, and much efforts should e taken to improve the environmental efficiency. Tale 3. Efficiency results Regions Model (2) Model (8) SO 2 Soot Dust Overall SO 2 Soot Dust Overall Beijing Tianjin Heei Liaoning Shanghai Jiangsu Zhejiang Fujian Shandong Guangdong Hainan Shanxi Inner M Jilin Heilongjiang Anhui Jiangxi Henan Huei Hunan Guangxi Chongqing Sichuan Guizhou Yunnan Shaanxi Gansu Qinghai Ningxia Xinjiang Mean We can find in Tale 3 that, if the learning effects are considered, nine regions are measured as efficient in the industrial systems, while other regions are all inefficient. Note that, once a regional industrial system is efficient under model (8), it is efficient in all waste gas emissions 19

20 performance. Inefficiency in each waste gas emission might e affected y the learning effects. Specifically, Beijing is rated as efficient under model (2), while is inefficient under the proposed model. This result means that, Beijing needs not to improve its environmental efficiency under model (2). However, under the proposed model, its efficiency scores of the three waste gas emissions are all less than unit, i.e., , and for SO 2, soot and dust, respectively. This indicates that some effective actions should e taken to reduce all the three waste gas emissions in Beijing under model (8). Interestingly, Jilin and Heilongjiang are inefficient under model (2) ut efficient under our proposed model. This implies that, these two regions need not to enhance their environmental efficiency under our proposed model ut should do under the traditional model. It is noteworthy that, Ningxia is ranked at the ottom of the overall efficiency scores (0.2734) and Guizhou has the second lowest overall efficiency score (0.2775). These results are confirmed y their efficiency scores in waste gas emissions. For example, Ningxia is evaluated to have the lowest dust emission efficiency score (0.0879) among all the regions and the second lowest efficiency scores of SO 2 and soot emissions, i.e., and , respectively. Note that, total emissions of SO 2, soot and dust (including the increases caused y the learning effects) from Ningxia s industrial system in 2010 are ton, ton and ton respectively, which are much larger than those of Qinghai s, i.e., ton, ton and ton, respectively. However, its GDP is illion RMB Yuan, which is only a little higher than that of Qinghai s (61.37 illion RMB Yuan). Based on the efficiency results as reported in Tale 3, the efficiency changes caused y the learning effects can e clearly oserved. For ease of analysis, we take the three areas as an example to illustrate this. The results are displayed in Tale 4. Tale 4. Efficiency changes Area SO 2 efficiency Soot efficiency Dust efficiency Eastern area Central area Western area China Note that in Tale 4, efficiency change for each waste gas emission is otained y the 20

21 efficiency under model (8) minus that under model (2). We can find that, the efficiency changes in SO 2, soot and dust of the eastern area are , and , respectively. These changes means that excluding the learning effects will reduce the efficiency scores of SO 2 and soot, while increase dust efficiency in the eastern area. This in turn indicates that the learning effects have positive impacts on SO 2 and soot efficiencies while negative on dust efficiency in this area. Oviously, the efficiency changes in the three waste gas emissions in the central area are all positive, which suggests that the learning effects in this area show negative influences on the performance of all waste gas emissions. As for the western area, the learning effects show positive impacts on soot efficiency while negative on SO 2 and dust efficiencies. These results suggest that, similar to the effects on the reductions in the three waste gas emissions as shown in Tale 2, there also exist evident disparities etween the efficiency changes among areas. Similar disparities can also e oserved among regions as shown in Tale 3. It can e oserved in Tale 4 that, the efficiency changes in all waste gas emissions of the whole country are all positive, and thus in turn means that the learning effects have significantly negative impact on the environmental efficiency for the whole national industrial system. These results might not e caused y the learning effects themselves during the study time period ut other production factors such as significantly increased laor and capital inputs or unalanced GDP outputs, or much more inputs to the learning activities. Although the environmental efficiency is worse off, the total reductions in the three waste gas emissions are huge. For example, total reduction in SO 2 emission caused y the learning effects for the whole national industrial system is ton, which is aout 3.66 times that in 2010 ( ton). To further illustrate the learning effects on the environmental efficiency, we next examine the reduction potentials of the three waste gas emissions otained under model (2) and model (8). To accurately calculate the reduction potentials for the waste gas emissions in 2010, we should exclude the impacts of the learning effects on reduction potentials of the waste gas emissions. To this end, we use DGPl 0 GPl 0 GP l0 to measure the difference etween reduction potentials for the waste gas emissions under model (8) and model (2). Note that, the target waste gas emissions under model (8) is yl 0 yl 0 GPl 0 ( l 1,..., p ). Then, ( yl 0 yl 0 GPl 0 ) yl 0= yl 0 GP l0 can e used to measure the 21

22 corresponding target emissions y excluding the learning effects. Thus, we have y ( y GP ) GP, which can e used to measure the real reduction potentials for the l0 l0 l0 l0 three waste gas emissions in 2010 with the consideration of the learning effects. In such a circumstance, DGP l0 can also e used to characterize the environmental efficiency changes for the waste gas emissions. Therefore, we have following remark: Remark 1: The learning effects on the environmental efficiencies follow the rules: (a) If DGP l0 0, the learning effects might have not significant impact on the environmental efficiency improvement for the waste gas emission l; () If DGP l0 0, the learning effects might have significantly positive impact on the environmental efficiency improvement for the waste gas emission l; (c) If DGP l0 0, the learning effects might have significantly negative impact on the environmental efficiency improvement for the waste gas emission l. Based on model (2) and model (8), the reduction potentials of the waste gas emissions and their difference can e otained. The results are shown in Tale 5. Note that in Tale 5, seven regions are estimate to have DGP 0 in SO 2, i.e., Tianjin, Shanghai, Fujian, Guangdong, Inner Mongolia, Henan and Chongqing. This means that, the learning effects have no significant impacts on their SO 2 emission efficiency improvements. The SO 2 reduction potentials as shown in Tale 5 under model (2) should e conducted y enhancing the performance of other production factors such as laor and capital input performance, GDP output performance rather than the learning performance. 19 regions such as Beijing, Huei, Liaoning and Zhejiang are estimated to have DGP 0 in SO 2, which indicates that these positive differences can e reduced y the cumulative learning effects. The larger the DGP, the more improvement the cumulative learning performance; and vice versa. For example, Shandong has the largest DGP (i.e., ). This suggests that Shandong should take much more efforts to improve its learning effects to reduce this difference, while the remainder reduction potential under model (2) should e decreased y enhancing other production factors operational performance. We also find that, four regions (i.e., Jilin, Heilongjiang, Qinghai and Xinjiang) are 22

23 estimated to have negative differences in SO 2, i.e., DGP 0, which implies that, the learning effects might have negative impacts on their improvements of the environmental efficiencies. It is cautious for these four regions to take some efforts to conduct the learning activities, ut to the enhancement of other factors operational performance. Similar oservations can also e found in the reduction potentials of soot and dust emissions among regional industrial systems. Tale 5. Reduction potentials a Regions SO 2 (10 4 ton) Soot (10 4 ton) Dust (10 4 ton) M 2 M 8 DGP M 2 M 8 DGP M 2 M 8 DGP Beijing Tianjin Heei Liaoning Shanghai Jiangsu Zhejiang Fujian Shandong Guangdong Hainan Shanxi Inner M Jilin Heilongjiang Anhui Jiangxi Henan Huei Hunan Guangxi Chongqing Sichuan Guizhou Yunnan Shaanxi Gansu Qinghai Ningxia Xinjiang a Note: M 2 and M 8 refer to model (2) and model (8), respectively. 23

24 It should e stated that, a region with DGP 0 does not mean that the learning effect has no impact on its industrial system s environmental efficiency improvement ut there is no significant improvement potential relative to the existing reference regions in the industrial systems. Not that, DGP results of Jilin and Heilongjiang are estimated to have negative values for all the three waste gas emissions as shown in Tale 5. These results suggest that, these two regions might have spent much more on the learning activities, thus in turn the learning activities lead to negative impacts on the whole environmental efficiency of the three waste gas emissions. These findings suggest that our proposed approach comined with traditional approach can effectively help to identify whether the environmental efficiency can e improved y enhancing the learning performance or other factors operational performance. These results provide important insights for the managers to manage their production operational activities and learning activities so as to effectively improve the overall environmental efficiency for the whole national industrial systems. 4. Conclusions We in this paper aim to examine the environmental efficiency of industrial systems in China y considering the learning effects. To this end, a SBM model y incorporating the learning effects is proposed. In the descried approach, the learning effects on waste gas emissions reductions are estimated ased on the log-linear learning curve model. The proposed approach can effectively help to identify the main causes of inefficiency that arise from the learning performance, which cannot e estimated y only using conventional DEA models. The proposed approach is then applied to examine the environmental efficiency of regional industrial systems in China with 2010 dataset. Results of the application show the following findings and conclusions: First, the overall environmental efficiency in China s industrial system is not relatively high. Second, evident disparities can e found in the learning effects on waste gas emissions reductions as well as the environmental efficiencies among regions and areas. Third, the learning effects have significant impacts on regional industrial systems environmental efficiencies as well as the reduction potentials of the waste gas emissions. Finally, an important management insight can e achieved. That is, the learning effects might have no significant influences on the environmental efficiency 24