Does energy efficiency label alter consumers purchase decision? A latent class approach on Shanghai data. May 11, Junyi Shen.

Transcription

1 OSIPP Dicuion Paper : DP-2007-E-005 Doe energy efficiency label alter conumer purchae deciion? A latent cla approach on Shanghai data May 11, 2007 Junyi Shen Aitant Profeor, OSIPP, Oaka Univerity and Tatuyohi Saijo Profeor, ISER, Oaka Univerity C25, C93, D12, Q49, Energy efficiency label, Conumer purchae deciion, Latent cla model, Willingne to pay, China In thi paper we apply hypothetical choice experiment through a field urvey in Shanghai of China to examine whether China Energy Efficiency Label affect conumer choice of air conditioner and refrigerator. A latent cla approach i ued to oberve both heterogeneitie among the repondent and product brand. The reult ugget that the effect of energy efficiency label on conumer preference i twofold. Firt, more energy efficient air conditioner or refrigerator are preferred by conumer, no matter whether they are with foreign brand or dometic brand and whether they are new or econd-hand. Second, energy efficiency label per e i recognized by conumer. In addition, preence of a (hypothetical) label that indicate the electricity bill difference comparing to a tandard model i ignificantly preferred by the repondent in mot of the cae, uggeting that more information provided to conumer make them much happier. Finally, the cla probability weighted willingne to pay value for one rank upgrading in energy efficiency of refrigerator are higher than thoe of air conditioner, implying that conumer have an incentive to pay more for appliance ued more frequently. The author would like to appreciate the financial upport of Japanee Minitry of the Environment (Global Environmental Reearch Fund H-062) and Japanee Minitry of Education, Culture, Sport, Science and Technology (Grant-in-Aid for Young Scientit (B), ). Comment and uggetion on the urvey iue from Yohiyuki Shimoda, Oamu Saito, Yugo Yamamoto, and Kenju Akai are greatly appreciated. All the view expreed in thi paper and any error are the ole reponibility of the author. Correpondence to: Junyi Shen, Oaka School of International Public Policy (OSIPP), Oaka Univerity, 1-31 Machikaneyama, Toyonaka, Oaka , Japan. junyi@oipp.oaka-u.ac.jp 1

2 1. Introduction Studie on conumer behavior are uually conducted by policy deciion maker for olution to environmental policy problem, becaue they reaon, at leat implicitly, that environmental degradation come from economic activity, that conumer expenditure account for mot of gro dometic product, and that therefore changing conumer behavior can go a long way toward reducing environmental degradation (Stern, 1999). During the lat decade, there ha been a number of public policie interet in utainable conumption and the role of conumer in environmental conervation. However, the needed change in conumption pattern toward utainable conumption normally cannot be achieved by force, but only with a kind of voluntary approach uch a information providing policy, which play a key role in environmental policy targeting conumer. One of the mot promiing form of environmental information policy, in term of providing timely and relevant information for the conumer, i environmental labeling that i authenticated and monitored by a third-party (Crepi and Marette, 2005; Loureiro et al., 2002; Thøgeren, 2000). The main purpoe of providing variou environmental label (alo called eco-label) to conumer i to avoid the aymmetric information problem between producer and conumer, becaue ome environmentally friendly product normally have unobervable characteritic, which will caue potential inefficiencie reulting from imperfect information. The idea that information proviion could be an effective method of environmental regulation i a ignificant and even revolutionary departure from traditional thinking (Ruell et al., 2005). Thi departure i explicitly recognized a the third wave in pollution control policy in Tietenberg and Wheeler (2001). 1 Beginning with the German Blue Angel in 1977, a number of environmental labeling program in the organic food and electricity market a well a the foretry product have developed (e.g. Nordic Swan in the Nordic countrie, EU Flower in EU countrie, Green Seal in USA, Environmental Choice in Canada, Eco-Mark in Japan, Green Label in Singapore, etc.). Encouraged by widely increaing environmental labeling cheme, both theoretical and empirical tudie in academic ociety on thi iue have been rapidly growing (e.g. Amacher et al., 2004; Banerjee and Solomon, 2003; Bjørner et al. 2004; Caon and Gangadharan, 2002; Crepi and Marette, 2005; Doi and Moretto, 2001; Howarth et al., 2000; Johnton et al., 2001; Loureiro et al., 2002; Moon et al. 2002; Morri et al., 1995; Roe et al., 2001; Ruell et al. 2005; Thøgeren, 2000; Tiel et al., 2002, etc.). Concerning the effect of environmental label on conumer purchae deciion, Bjørner et al. (2004) 1 The firt wave refer to a command-and-control regulation, while the econd wave involve the introduction of attempt to alter polluter behavior through the ue of economic incentive, or what are commonly called market-baed intrument uch a emiion charge, tradable dicharge permit, and depoit-refund cheme (Ruell et al., 2005). 2

3 ued a large Danih conumer panel with detailed information on actual purchae from the beginning of 1997 to January 2001 to quantify the effect of the Nordic Swan on conumer brand choice of toilet paper, paper towel and detergent. They found that the Nordic Swan label ha had a ignificant effect on conumer brand choice for toilet paper and detergent. Similar effect of variou eco-label on conumer choice of product can alo be found in the tudie of dolphin-afe label (Teil et al., 2002), eco-labeled apple (Loureiro et al., 2002), eco-labeled eafood (Johnton et al., 2001), eco-labeled agricultural product (Moon et al., 2002), and energy label (Banerjee and Solomon, 2003; Howarth et al., 2000). In addition, two theoretical tudie (i.e. Amacher et al., 2004; Doi and Moretto, 2001) tated that eco-labeling could be ued a a mean of reliable environmental policy to improve market outcome by which firm adopt clean(er) technologie and produce green product baed on conumer recognition on eco-label and willingne to pay for thee certified environmentally friendly product. Furthermore, Caon and Gangadharan (2002) examined variou treatment that could remedy the market failure ariing from incomplete information through laboratory poted offer market. They concluded that the only reliable way to improve product quality in the experiment i to ue a third party that charge a fee to certify product quality, i.e., environmental labeling. To provide more empirical evidence, in thi tudy we preent an analyi of the effect of a pecific environmental label (China Energy Efficiency Label) on Chinee conumer choice among different brand of air conditioner and refrigerator. Nowaday, energy label and minimum energy performance tandard are fat becoming commonplace throughout the world. It i axiomatic that the market for houehold energy ervice would be enhanced where buyer are able to take into account not jut the cot of the appliance but the otherwie inviible factor of energy conumption. Energy label improve the market operation by diplaying accurate energy conumption information on product, which i thought to be ueful in the purchae deciion. There are two main type of energy label: endorement and comparative label. Endorement label indicate that product belong to the mot energy efficient cla of product or meet a predetermined tandard or eligibility criteria. Comparative label uch a China Energy Efficiency Label allow conumer to form a judgment about the energy efficiency and relative ranking of all product that carry a label. Endorement and comparative label can coexit, and do o in many countrie. The main difference between endorement label and comparative label i that the effect of the latter on conumer purchae deciion i twofold: (i) whether or not the label per e i recognized by conumer; (ii) whether or not different rank preented on the label, which indicate different level in energy efficiency, alter conumer preference and purchae deciion. Previou work on the effect of energy label on conumer repone i limited. Banerjee and Solomon (2003) preented a meta-evaluation of five US energy labeling 3

4 program: Green Seal, Scientific Certification Sytem, Energy Guide, Energy Star, and Green-e. They tated that although mot people indicate a trong preference for a credible, third-party labeling ytem and claim to ue it, very few actually ue the label for purchaing deciion and Such overtated repone leading to urvey bia are thought to be due to the warm glow effect. Acknowledging thi kind of warm glow effect and to overcome it in our tudy, we apply a choice experiment approach in which repondent are aked to elect a mot preferable alternative from a et of alternative baed on a number of attribute and their level. Since the repondent choice deciion i made baed on the tradeoff among the aociated attribute, we believe that warm glow effect, at leat to ome extent, could be avoided by thi manipulation. The analyi i baed on hypothetical choice experiment for air conditioner and refrigerator through a quetionnaire urvey conducted at the beginning of November 2006 in Shanghai of China. In each choice experiment, we provide four alternative named a new product A with foreign brand, econd-hand product B with foreign brand, new product C with dometic brand, and econd-hand product D with dometic brand, repectively. 2 Thi pecification allow u to oberve Shanghai conumer preference on econd-hand product and different brand (foreign or dometic). In addition to the attribute aociated with energy efficiency label, other attribute include the price, the electricity conumption, applicable pace/volume, with or without air cleaning/ilence function, and preence or abence of a label indicating the electricity bill difference comparing to a tandard model. Note that including the final attribute related to preence or abence of another label erve a a econd goal of thi tudy to invetigate whether or not additional information proviion affect conumer purchae deciion. A Latent Cla(LC) model approach i applied in the etimation. Compared to a tandard Multinomial Logit (MNL) model in dicrete choice, the LC approach allow the analyt to oberve individual heterogeneity through identifying and characterizing variou preference group (Boxall and Adamowicz, 2002; Louviere et al., 2000; Greene and Henher, 2003; Shen, 2006). In addition, thi approach can etimate imultaneouly the probability of individual in each cla, hence, it i even valid in welfare analyi which conider to aggregate welfare gained from each cla. A a preview of reult, it i found that China Energy Efficiency Label doe have a ignificant effect on Shanghai conumer purchae deciion of air conditioner and refrigerator. Meanwhile, higher price and higher electricity conumption reduce individual preference. Function uch a air cleaning of air conditioner and ilence of refrigerator are preferred by conumer in mot cae, while applicable pace of air conditioner and volume of refrigerator are evaluated differently by the repondent. 2 The fact that there exit a econd-hand appliance market in China encourage u to examine the iue on Chinee conumer preference of econd-hand appliance. 4

5 Furthermore, conumer prefer to have a (hypothetical) label indicating clearly the electricity bill difference comparing to a tandard model in mot cae, uggeting that information proviion at an explicit way uch a through a label i an effective mean to affect conumer choice deciion. The remainder of the paper i organized a follow. Section 2 provide a brief ummary of China Energy Efficiency Label program. Section 3 preent the model pecification. The urvey iue i decribed in Section 4 and the empirical reult are reported in Section 5. Finally, we provide concluion remark and future implication in Section China Energy Efficiency Label program Home appliance ownerhip and production ha increaed dramatically in China in the pat two decade. From extremely low level in 1980, China appliance indutry ha become one of the larget in the world, with ale topping U.S.$ billion in Meanwhile, the ownerhip ratio of refrigerator by urban Chinee houehold increae from le than 1 percent in 1981 to over 90 percent by 2006 (China State Statitic Bureau, 2006). A imilar trend ha been oberved for room air conditioner and other home appliance a well. Such a dramatic increae in appliance ownerhip ha ignificantly led to the growth of China electricity ue. Between 1980 and 2001, per capita annual electricity conumption grew from to kwh, with an average annual growth rate of 8 percent, while per capita reidential electricity ue grew from 10.0 to kwh, averaging 15 percent growth per year (nearly twice a fat a overall electricity conumption). The growth in houehold electricity conumption ha contributed ubtantially to the tremendou increae of generating capacity in China in recent year. Since 1990, China ha added on average 16 GW of new capacity each year. Well acknowledging the environmental impact of electricity generation, China tarted to implement a number of program including energy labeling program to improve energy efficiency. 3 The bai for energy labeling program in China wa etablihed in the Energy Conervation Law, which wa approved by the National People Congre on November 1, 1997 and came into force on January 1, The law upereded earlier law and et the bai for meaure to develop energy efficiency tandard and energy labeling of appliance and equipment. At firt, the government decided to iue a voluntary endorement label China Energy Conervation Label (ee Figure 1). The State Economic and Trade Commiion 3 Gain in energy efficiency ha been regarded a one of the explanation for China decline in energy intenity in recent year. For detail on thi iue, ee Fiher-Vanden et al. (2004). 5

6 (SETC, now NDRC: National Development and Reform Commiion) and the China State Bureau of Quality and Technical Superviion (CSBTS, now AQSIO: Adminitration for Quality, Superviion, Inpection and Quarantine) together etablihed an independent, non-profit organization the China Certification Center for Energy Conervation Product (CECP), aiming at developing requirement in order to certify product a being afe, high-quality, and energy-aving. The endorement label deigned by the CECP wa formally launched in September 1999 and firt applied in refrigerator. The application for China Energy Conervation Label i imilar to the USA Energy Star Label and o far ha been awarded to over 100 of product. In addition to the above voluntary energy conervation label, the China National Intitution of Standardization (CNIS) wa aigned to develop a mandatory and comparative energy information label China Energy Efficiency Label (ee Figure 2). The deigned label came into effect from March The firt two product were refrigerator and room air conditioner. The China Energy Efficiency Label i imilar to EU Energy Label, which rate the energy efficiency of the appliance in term of a et of energy efficiency rank from 1 to 5 on the label. Rank 1 i the mot energy efficient, while rank 5 i the leat efficient. The CNIS report that by the end of March 2006, the product with energy efficiency rank 1 and 2 in China hare about 10 percent in room air conditioner market and 68 percent in refrigerator market. In contrat, the product with energy efficiency rank 5 only occupy 7 percent in room air conditioner market and 8 percent in refrigerator market. Thee evidence ugget that there i an energy efficiency rank gap between air conditioner and refrigerator in China. Compared to the majority of high efficient energy-aving refrigerator on the market, mot of room air conditioner are produced with medium energy efficiency level (rank 3 and 4). 3. Model pecification Choice model i baed on random utility theory. The baic aumption embodied in the random utility approach to choice modeling i that deciion maker are utility maximizer, i.e., given a et of alternative the deciion maker will chooe the alternative that maximize her utility. The utility of an alternative for an individual (U ) i modeled to conit of a determinitic component (V ) and a random error term (ε ). Formally, individual q utility of alternative i can be expree a: U = + ε (1) iq V iq iq Hence the probability that individual q chooe alternative i from a particular et J that comprie j alternative can be written a: P iq = P( U > U ; j( i) J ) = P( ε < ε + V V ; j( i) J ) iq jq jq iq iq jq (2) 6

7 To tranform the random utility model into a choice model, certain aumption about the joint ditribution of the vector of random error term i required. If the random error term are aumed to follow the type I extreme value (EV1) ditribution and be independently and identically ditributed (IID) acro alternative and cae (or obervation), the multinomial (or ometime called conditional) logit (MNL) model (McFadden, 1974) i obtained. In the MNL model, the choice probability in Equation (2) i expreed a: P iq = exp( µ V ) / exp( µ V iq J j= 1 jq ) (3) Then, making further aumption for the determinitic component of utility to be linear and additive in parameter, Viq = β Xiq, the probability in Equation (3) can be given a: P iq = exp( µβ X iq ) / J j= 1 exp( µ β X jq ) (4) where µ repreent a cale parameter that determine the cale of the utilitie, which i typically normalized to 1.0 in the MNL model. Xiq are explanatory variable of Viq, normally including alternative-pecific contant (ASC), the attribute of the alternative i and the ocial-economic characteritic of the individual q, β i the parameter vector aociated with the matrix Xiq. It i well known that heterogeneity among individual i extremely difficult to examine in the MNL model. Thi limitation could be relaxed, to ome extent, by interacting individual-pecific characteritic with variou of the choice. However, thi method i limited becaue it require a priori election of key individual characteritic and attribute and only involved a limited election of individual pecific variable (Boxall and Adamowicz, 2002). One way of circumventing thi difficulty i through etimating the model by the Latent Cla (LC) model. The LC model aume that the population conit of a number of latent clae S and the unoberved heterogeneity among individual can be captured by thee clae through etimating a different parameter vector in the correponding utility function. Formally, the choice probability of individual q of cla i expreed a: P J = iq exp( β X iq ) / exp( µ β X jq ) j= 1 µ 1,..., S = (5) where µ and β are cla-pecific cale and utility parameter, repectively. Then, following Boxall and Adamowicz (2002), Louviere et al. (2000), and Swait (1994, 2007), the probability of individual q in cla (Hq ) can be expreed a: H = exp( αλ Z ) / exp( αλ Z ) q q q = 1 S (6) 7

8 where α i a cale factor normally normalized to 1.0, λ i the parameter vector in cla, and Z denote a et of characteritic (e.g. individual-pecific characteritic) q determining the claification probability. Combining conditional choice equation (5) and memberhip claification equation (6), the unconditional probability of chooing alternative i i given a: S S J S Piq = Piq Hq = exp( µβ Xiq ) / exp( µβ X jq ) exp( αλ Zq ) / exp( αλ Zq ) = 1 = 1 j= 1 = 1 (7) In equation (7), when we et µ and α equal to one 4, the parameter vector β and λ can be imultaneouly etimated by the maximum likelihood method to explain choice behavior. 5 However, the LC model cannot be etimated unle S (the number of clae) in equation (7) i given, becaue S i dicrete but maximum likelihood etimation theory require that the parameter pace be continuou and etimate be in the interior of the pace (Swait, 2007). Therefore, the central iue in the LC model i how to determine S. The literature ha recommended a number of information criteria to determine S (e.g. Boxall and Adamowicz, 2002; Greene and Henher, 2003; Louviere et al., 2000; Morey el al., 2006; Shen, 2006; Swait, 2007, etc.). Among them, four meaure baed on the log likelihood at convergence with clae, ample ize and number of parameter are popular to be ued to determine S. They are defined a: Akaike Information Criterion, AIC = 2(log L* K ) (8) Akaike ρ 2, ρ 2 = 1 [ AIC /(2 log L )] (9) 0 Bozdogan Akaike Information Criterion, AIC3= 2logL* + 3K (10) Bayeian Information Criterion, BIC = log L* + ( K log N) / 2 (11) where log L* i the log likelihood at convergence with clae, K i the number of parameter in the model with clae, i the log likelihood of the ample with equal choice probabilitie, and N i the ample ize. L 0 An alternative approach accounting for individual heterogeneity i called Random Parameter Logit (RPL) or Mixed Logit (ML) model, which allow model parameter to vary randomly through aumed ditribution (e.g. normal, log-normal, triangular, etc.) over individual (e.g. Bhat and Goen, 2004; Bjørner et al., 2004; Greene and Henher, 2003; He et al., 2005; McFadden and Train, 2000; Revelt and Train, 1998; Train, 1998, etc.). In thi approach, each individual ha their own et of cale and utility parameter. 4 Boxall and Adamowicz (2002) note that utilizing the LC model in empirical etimation require that all the cale factor in equation (7) are et equal to one. 5 The parameter vector λ in one of the latent clae mut be normalized to zero (e.g. λ 1 0 ) to run the etimation. Therefore, the remaining λ are identified relative to thi normalization. 8

9 From thi viewpoint, one could regard the RPL/ML model a the cae where each individual in the ample could be conidered a an individual cla, which i indeed the LC model with N (ample ize) clae. In other word, the LC model control individual heterogeneity with clae where i between 1 and N. Compared to the RPL model, the potential advantage of the LC model are twofold. Firt, the LC approach i emiparametric, therefore, it doe not require the analyt to make pecific aumption about the ditribution of parameter acro individual (Greene and Henher, 2003). Second, the LC model yield the probabilitie in each cla. Thi mean that although each repondent i aumed to belong to one cla, it i taken into account that there i uncertainty about a repondent cla memberhip. 4. Survey iue 4.1 Choice experiment deign In the choice experiment, a number of attribute and aigned level are ued to generate hypothetical cenario. The attribute and their level included in each cenario for thi tudy are ummarized in Table 1 (air conditioner) and 2 (refrigerator). We have four alternative (product) named a new product A with foreign brand, econd-hand product B with foreign brand, new product C with dometic brand, and econd-hand product D with dometic brand, repectively for both air conditioner and refrigerator. 6 Each air conditioner or refrigerator ha ix attribute. For air conditioner, they are (i) price (three different level for each alternative) (ii) hourly electricity conumption (two ame level for A and C, and two ame level for B and D) (iii) applicable pace (two level for all alternative) (iv) whether there i an air cleaning function (two level for all alternative) (v) energy efficiency rank preented on the label (four level for all alternative) (vi) whether there i a label of indicating the electricity bill difference comparing to a tandard model (two level for all alternative). 7 6 A non-choice (the choice not to elect one of the available alternative) alternative could be conidered to provide. However, following the argument of Henher et al. (2005) that if the objective of the tudy i to examine the impact of the relationhip different attribute level have upon choice, then any non-choice alternative i likely to be a hindrance to the analyt, we did not provide a non-choice alternative in thi tudy. 7 Including the label of indicating the electricity bill difference comparing to a tandard model a an attribute of appliance i to examine in addition to energy efficiency label whether information proviion about running cot aving through a label could be an effective iue influencing conumer purchaing deciion. Thi could 9

10 For refrigerator, three attribute are the ame a (i), (v), and (vi) for air conditioner, another three attribute are conidered a daily electricity conumption (two ame level for A and C, and two ame level for B and D), volume (two level for all alternative), and whether there i a ilent function (two level for all alternative). Choice experiment deign i concerned with how to create the choice et in an efficient way, i.e. how to combine attribute level into profile of alternative and profile into choice et (Alpízar et al., 2003). It i obviou that it would generate too many choice et if we apply a full factorial deign, which i too much for repondent to anwer even we divide the choice et into a number of verion. In thi tudy, we adopt the D-optimal deign approach for choice experiment baed on multinomial logit model. The objective of the D-optimal deign i to extract the maximum amount of information from the repondent ubject to the number of attribute, attribute level and other characteritic of the urvey uch a cot and length of the urvey (Carlon and Martinon, 2003). The D-optimal deign i implemented to maximize a choen optimality criterion (e.g. D-, A-, and G-efficiency) baed on the pre-pecified model (e.g. multinomial logit model in a choice experiment). One common meaure of efficiency i called D-efficiency that i applied in thi tudy and given a 1 1/ K D-efficiency = Ω (12) where K i the number of parameter to etimate, Ω i the covariance matrix of a vector of parameter. Beide D-efficiency, there are alo everal other criteria of efficiency uch a A- and G-efficiency. The main reaon for chooing D-efficiency i that it i le computationally burdenome and could be directly run by a number of tatitical oftware. 8 A a reult of running the D-optimal deign through Deign-Expert 7.0 (Stat-Eae, Inc.), we created 48 choice et for air conditioner and 48 choice et for refrigerator, repectively. Thee choice et were further randomly divided into 8 verion, i.e., each verion of the quetionnaire conit of 6 choice et for air conditioner and 6 choice et for refrigerator. The repondent were aked to elect the mot favorite air conditioner or refrigerator in each choice et and anwer everal other quetion related to their ocioeconomic characteritic. The example of choice et for air conditioner and refrigerator are preented in Table 3 and 4, repectively. 4.2 Data collection A urvey tudy wa conducted at the beginning of November 2006 in Shanghai of China. be viewed a an additional tet for the validity of information proviion through label program. 8 For more detail on D-optimal deign, ee, for example, Alpízar et al. (2003), Carlon and Martinonb (2003), and Huber and Zwerina (1996), etc. 10

11 Thi urvey aimed to evaluate Shanghai reident preference and awarene on China Energy Efficiency Label, which i a mandatory and comparative energy information label in China (ee detailed introduction in Section 2). The urvey wa conducted by two profeional marketing firm. One firm called Nikkei Reearch wa in charge of collecting 600 obervation through face-to-face interview. The repondent were randomly recruited on the treet of two ditrict (buine center ditrict and reidential ditrict), with 300 obervation in each ditrict. The trained invetigator were aigned to ak the repondent variou quetion according to the quetionnaire. The average time for the face-to-face interview wa approximate 15 minute. In addition, another firm called Searchina Reearch conducted the web-baed urvey and alo collected 600 valid ample. The quetionnaire for both urvey were with the ame content, which include (i) a number of quetion to reveal the repondent environmental concern 9 ; (ii) choice of air conditioner and refrigerator; and (iii) repondent ocioeconomic characteritic uch a gender, age, education level, occupation, annual houehold income, and houehold ize, etc. The ummary of ocioeconomic characteritic for both ample i provided in Appendix 1. From the ummary, we may find out that both ample cover mot type of the reident and houehold in Shanghai. In addition, compared to face-to-face interview ample, web-baed ample eem encounter a potential bia in age and education level. 5. Empirical reult Table 5-13 preent the reult aociated with the MNL/LC pecification and WTP and elaticity etimate. The reult preented were analyzed by uing NLOGIT 3.0, a pecialit dicrete choice modeling package in LIMDEP (Econometric Software, Inc.). The definition of the variable ued in thi tudy can be found in Appendix 2. A a whole impreion of the MNL and LC etimate, we may find that compared to the MNL model, the goodne of fit meaure (Peudo R 2 and predictive power) get ignificantly improved after applying the LC approach. 5.1 Likelihood Ratio tet for ample election We tart by dicuing the reult of Likelihood Ratio (LR) tet for whether or not the two ample (face-to-face interview, hereinafter interview, and web-baed urvey, hereinafter web) could be pooled. It i aid that if the etimated utility parameter are 9 Thee quetion aiming at revealing the repondent environmental concern are for the purpoe of another tudy, which examine the ocioeconomic determinant of individual environmental concern. See Shen and Saijo (2007) for detail. 11

12 equal acro interview ample, web ample, and the pooled ample, then we may etimate the model by pooling thee two ample. The null hypothei on equal utility parameter i formally given a H0: interview web interview+web β = β = β The tatitic of the likelihood ratio tet uggeted by Swait and Louviere (1993) are calculated a: for air conditioner model, LR=-2( ( ))=167.76; for refrigerator model, LR=-2( ( ))= Since the critical value of the chi-quare i at the 1% ignificance level on 21 degree of freedom, the null hypothei that the vector of common utility parameter i equal acro two ample can be rejected in both air conditioner and refrigerator cae. Therefore, the following etimation are conducted by each ample. 5.2 Determining the number of latent clae A dicued in Section 3, the meaure of AIC, ρ 2, AIC3 and BIC were applied to help determining the number of latent clae. We attempted variou number of clae (1, 2, 3, 4 and 5 clae) and ummarize the tatitic in Table 5. The log likelihood value at convergence reveal that in all cae the model fit improve with the number of clae, epecially with the 2- and 3-cla model. Thi i not urpriing becaue log likelihood value normally increae in magnitude when there are more parameter need to be etimated. From the meaure of AIC3 and BIC (column 6 and 7), we may find that the minimum value of AIC3 and BIC are clearly related to the 3-cla model in all cae, uggeting that 3-cla model i optimal. Furthermore, the minimum AIC and the maximum 2 ρ eem alo upport 3-cla model a the bet olution in mot cae. There i an exception in refrigerator model of interview ample, which ugget that 4-cla model i better than 3-cla one. However, it i noteworthy that in thi cae the improvement from 3-cla to 4-cla i negligible becaue the reduction of AIC and increae of 2 ρ are o mall. Baed on the above dicuion, we therefore determined to elect 3-cla for etimating the LC model in thi tudy. 5.3 Characterizing the cla member Reult of cla memberhip for the 3-cla LC model in air conditioner and refrigerator are reported in Table 6 and 7. We claified the 3 latent clae of the repondent by their ocioeconomic characteritic. 10 Note that the variable aociated with houehold 10 An argument can be made for including the repondent environmental concern core in the cla memberhip claification model, ince the repondent environmental concern could alo be aociated with the claification of variou cla memberhip. However, a we analyzed in another tudy (Shen and Saijo, 2007), we found that the 12

13 ize and occupation tatu are not lited in the table, due to the reaon that thee two ocioeconomic characteritic are not ignificant even at 10% ignificance level in any cae. 11 The parameter for the third cla in both table (column 4 and 7) are et to 0 due to their normalization during etimation. Therefore, the parameter of the other two clae hould be explained relative to thi third cla. In air conditioner cae, the firt cla of interview ample could be claified a middle age (36-55 year old) or old age group (above 56 year old), or univerity/college graduate becaue the dummy variable of mid-age, old-age, and high-edu are ignificant at 5% ignificance level. We may find that the effect of age on determining cla memberhip i larger than that of education for thi cla, ince the etimated parameter of old-age and mid-age are about 1.3 time bigger than that of high-edu. For the econd cla, ince male i etimated with ignificantly negative ign and old-age i with ignificantly poitive ign, we label thi cla a old age group or female. Beide the labeled two clae, the remained repondent in interview ample are claified to be in the third cla. On the other hand, in web ample, cla 1 and cla 2 are characterized a middle age group or univerity/college graduate and high income group with annual houehold income more than 100 thouand Chinee yuan, repectively. 12 With repect to refrigerator cae, the claification of three clae for interview ample i imilar to that in air conditioner cae, with the difference that intead of old age group, high houehold income group i more likely to be in cla 2. For web ample, baed on the tatitical ignificance and ign of each variable, the firt cla i characterized a univerity/college graduate or income group with annual houehold income below 100 thouand yuan, while the econd cla i claified a income group with annual houehold income below 100 thouand yuan or male. 5.4 Reult of the 3-cla LC model The etimated reult of the 3-cla LC model are lited in Table For purpoe of comparion, the MNL etimate are alo provided in thee table. We etimate the alternative pecific contant (ASC) of air conditioner A, B, C in air conditioner model and refrigerator A, B, C in refrigerator model. Note again that for both appliance, A, B, ocioeconomic characteritic are highly correlated with each environmental concern meaure. Therefore, if we include thee meaure into claification model, it would caue eriou co-linearity among the independent variable. Hence, in thi tudy it i aumed that the repondent environmental concern meaure influence the cla memberhip through their ocioeconomic characteritic. We believe that thi hould be the iue. 11 There i no any ignificant effect on the model goodne of fit and other variable ignificance when we drop thee two characteritic from the model thouand yuan approximate thouand US$ if 1 US$=7.8 yuan. 13

14 and C tand for new product with foreign brand, econd-hand product with foreign brand, and new product with dometic brand, repectively. Beide the ASC, we etimate the alternative pecific attribute to examine the repondent preference on all the pecified attribute in detail. We believe that thi approach can allow u make the repondent preference on different alternative more clear. In other word, heterogeneity in alternative pecific attribute can be examined. Alternative pecific contant (ASC) Look at firt the parameter of three alternative pecific contant (A_ac, B_ac, and C_ac) in Table Mot of the ASC of A and C are etimated with ignificantly poitive ign, uggeting that individual prefer new air conditioner or refrigerator per e no matter it i a foreign or dometic product. In both interview and web ample, the A_ac of refrigerator in cla 3 and C_ac of air conditioner in cla 2 are not ignificant, implying that at leat ome repondent do not care whether it i a new product or not. Note that thee effect are abent in the ingle MNL model, exhibiting that the heterogeneity among individual could be captured by the latent cla approach. The reult of B_ac are omewhat mixed. In interview ample, the contant of air conditioner B i inignificant in all clae, while that of refrigerator B i ignificant in all clae. In particular, the B_ac of refrigerator in cla 1 i etimated with negative ign, uggeting that individual belonging to thi cla dilike ued refrigerator even it i with foreign brand. Thi kind of difavor with ued refrigerator or air conditioner can alo be found in web ample, for example, in cla 3 for refrigerator (ee B_ac in Table 11) and cla 2 for air conditioner (ee B_ac in Table 9). Conidering the probabilitie of repondent in thee clae, we may conclude that compared to ued eaonal appliance (i.e. air conditioner), more repondent dilike econd-hand daily appliance (i.e. refrigerator). Energy efficiency rank lited on energy label The key iue being worthy of remark in thi tudy i how individual evaluate the energy efficiency rank preented on energy label. From Table 8-11, it i noteworthy to ee that all the parameter aociated with energy efficiency rank are ignificant and with expected negative ign in all the clae of both ample. A we dicued in the Introduction ection, the effect of energy efficiency label on conumer preference i twofold, i.e. if the information provided by the mean of energy efficiency label work, conumer hould ignificantly react on the rank indicating energy efficiency on the label. Thu, the explanation on thee etimated ignificant and negative parameter can be made a (i) more energy efficient air conditioner and refrigerator are preferred, no matter whether they are new or econd-hand and whether they are with foreign or 14

15 dometic brand; (ii) an energy efficiency label per e i recognized by the conumer, otherwie conumer would not have ignificant effect of the energy efficiency rank on their preference. We argue that thi highly ignificant effect i mot poible due to the fact that in recent year individual environmental concioune and concern have been rapidly increaing in Shanghai. 13 To illutrate how individual evaluate the energy efficiency rank with monetary value, we provide the Willingne to Pay (WTP) value for both two appliance in Table 12. Conumer willingne to pay for one rank upgrading in energy efficiency i given by the equation: WTP a β ar = (13) β ap where β and ap β denote the etimated parameter aociated with attribute of ar price and energy efficiency rank in each cla for alternative a, repectively. Becaue thee two parameter vary acro clae and alternative, therefore, the etimated WTP value could identify heterogeneity among individual for the energy efficiency rank of different air conditioner and refrigerator. In addition, cla probability weighted WTP in Table 12 i calculated by applying the following equation: WTP a_ weighted a = 1 3 = WTP P (14) where P i the probabilitie of repondent in cla. Note that in everal cae, the etimated parameter of energy efficiency rank i not ignificant. We treat the WTP value aociated with thee inignificant parameter a zero. From table 12, we oberve everal evidence. Firt, the WTP value do vary acro the group, alternative, and appliance. Second, the etimated cla pecific WTP a 13 In the quetionnaire, we created nine quetion in order to reveal the repondent concern about general environmental iue and ome pecific environmental problem including global warming, cro-boundary and acid rain, air/water/oil pollution, urban energy problem, green land and ecological problem, the effect of harmful ubtance on health, dipoal/reduction/recycling of wate, and living environmental problem. The urveyed repondent were aked to chooe to what extent they concern about thee environmental iue. The percentage of anwer of concern and omewhat concern in interview and web ample are repectively 91% and 93.2% for general iue, 80.7% and 87% for global warming, 67.3% and 83.5% for cro-boundary and acid rain, 78.5% and 87.2% for air/water/oil pollution, 74.2% and 83.8% for urban energy problem, 83.3% and 91.2% for green land and ecological problem, 89.5% and 91.3% for the effect of harmful ubtance on health, 86.5% and 88.3% for dipoal/reduction/recycling of wate, and 93.8% and 91.2% for living environmental problem. See more detail in Shen and Saijo (2007). 15

16 well a cla probability weighted WTP value for refrigerator are larger than thoe correponding value for air conditioner in mot cae of both ample, uggeting that in contrat to an appliance ued eaonally, conumer are willing to pay more money for energy efficiency if the appliance i ued frequently. Similar effect ha been alo found in Bjørner et al. (2004), which tated that environmental labeling may be more effective on product purchaed more frequently, a conumer may feel that it would make a greater environmental impact. Third, the WTP value for econd-hand appliance B are lower than other two alternative A and C in each cla. However, the reult of comparion between A and C i mixed. In ome cae, the WTP value for A energy efficiency rank are higher than thoe for C, while in other cae they are revered. Taking into account the cla probability, we find that the cla probability weighted WTP value for C are larger than thoe for A. Thi evidence indicate that individual are willing to pay more on energy efficiency rank of thoe appliance with dometic brand. Finally, we provide direct elaticitie of energy efficiency rank on choice probability in Table 13. A hown in the table, all the elaticity value, although mall in magnitude, are negative to choice probability. Thi reult combined with the above dicuion are intuitive to both firm leader and government deciion maker becaue energy efficiency rank preented on the energy label i well recognized by the conumer and alter conumer purchae deciion. Therefore, China Energy Efficiency Labeling program could be regarded a effective baed on the criterion of conumer repone. 14 Other attribute effect Concerning the attribute of monetary cot i.e. price and electricity conumption that could be viewed a a kind of running cot, almot all the parameter aociated with A and C of both appliance are etimated with ignificant and negative to individual preference in both ample, which i conitent with economic theory a expected. For air conditioner B in interview ample, thee two parameter are not ignificant. We doubt that thi i mot probably due to the reaon that econd-hand air conditioner B ha no effect on conumer preference in thi ample (ee B_ac in Table 8). An air conditioner with air cleaning function or a refrigerator with ilent function i preferable in mot clae of both ample, no matter it i a new or econd-hand product. In addition, compared to thoe etimated parameter in web ample, more applicable pace of air conditioner almot doe not affect individual preference in interview 14 To formally evaluate a labeling program, it i neceary to conider both conumer repone and producer repone. However, examining the effect of energy efficiency label on firm production deciion i not the iue of the preent tudy and will be left a our future tak. 16

17 ample. Concerning the volume of refrigerator, high-capacity refrigerator raie the preference of thoe repondent belonging to clae 1 for all product in interview ample. However, for other cla member in interview ample and all the repondent in web ample, the preference of thi attribute differ according to different alternative. For example, in cla 2 of both ample, the repondent care about the volume of new refrigerator, while in cla 3 of web ample, the member prefer high-capacity refrigerator with foreign brand for both new and econd-hand one. Meanwhile, another attribute, which i preence or abence of a label indicating clearly the electricity bill difference comparing to a tandard model, i etimated with ignificant and poitive ign in mot cae epecially for new product A and C. Thi evidence ugget that individual would be more pleaed if more information i provided explicitly. Combined with the reult of energy efficiency label, it ugget that information proviion, at leat to ome extent, could be viewed a a valid mean to affect conumer preference and conequently their behavior. 6. Concluion In thi tudy we apply a urvey data et to etimate model for conumer choice among different brand of air conditioner and refrigerator, aiming at examining the effect of China Energy Efficiency Label on conumer preference. The latent cla approach allow u to control heterogeneity among individual that cannot be oberved through a imple multinomial logit model. Four alternative, which include two new one with foreign and dometic brand and two econd-hand one with foreign and dometic brand, are conidered in our choice experiment. Thi manipulation make it poible for u to invetigate conumer preference on econd-hand product and conequently their attribute. From the empirical analyi, it appear that energy efficiency rank preented on energy label have a ignificant effect on the choice of air conditioner and refrigerator, perhap becaue in recent year individual environmental concioune and concern have been rapidly increaing in Shanghai. The reult from the cla pecific willingne to pay for one rank upgrading in energy efficiency indicate that for econd-hand product the value are relatively lower than other two new product. Furthermore, the cla probability weighted willingne to pay value ugget that conumer prefer to pay more on product ued more frequently, a they probably feel that it would make a greater environmental impact and/or increae their running cot. However, examining more deeply to ditinguih between thee two poible reaon i impoible in the preent tudy. We leave thi iue a a future tak. Preence of another label indicating clearly the electricity bill difference comparing to a tand model i alo etimated to have ignificant effect on conumer 17

18 preference in mot cae. Thi evidence ugget that information provided explicitly through a label i preferred by conumer and could be conidered a a valid alternative to olve aymmetric information problem caued from market failure. The fact reulted from thi tudy that Shanghai conumer do react to the energy efficiency rank preented on China Energy Efficiency Label i intuitive to both firm leader and policy deciion maker. For the government, it i a great confidence to ay that environmental labeling i one of the effective policie to olve environmental iue. For the firm, it may create an incentive to invet in green technology or make them more confident in enlarging the cale of green production to meet conumer need. An unanwered quetion whether the effect of China Energy Efficiency Label on conumer purchae deciion found in Shanghai will alo be found in other province of China i remained. We leave it a an open quetion and encourage a great effort to tudy thi iue at much deeper extent. Reference Alpízar F., F. Carlon, and P. Martinon Uing choice experiment for non-market valuation. Economic Iue 8: Amacher, G. S., E. Kokela, and M. Ollikainen Environmental quality competition and eco-labeling. Journal of Environmental Economic and Management 47: Banerjee, A. and B. D. Solomon Eco-labeling for energy efficiency and utainability: a meta-evaluation of US program. Energy Policy. 31: Bhat C. R. and R. Goen A mixed multinomial logit model analyi of weekend recreational epiode type choice. Tranportation Reearch Part B 38: Bjørner, T. B., L. G. Hanen, and C. R. Ruell Environmental labeling and conumer choice an empirical analyi of the effect of the Nordic Swan. Journal of Environmental Economic and Management 47: Boxall, P. C. and W. L. Adamowicz Undertanding heterogeneou preference in random utility model: A latent cla approach. Environmental and Reource Economic 23: Carlon F. and P. Martinon Deign technique for tated preference method in health economic. Health Economic 12: Caon, T. N. and L. Gangadharan Environmental labeling and incomplete conumer information in laboratory market. Journal of Environmental Economic and Management 43: China State Statitic Bureau China Statitic Yearbook Crepi, J. M. and S. Marette Eco-labelling economic: i public involvement neceary?. in S. Krarup and C. S. Ruell (ed), Environment, Information and Conumer Behaviour. Edward Elgar Publihing. 18

19 Doi, C. and M. Moretto I ecolabelling a reliable environmental policy meaure?. Environmental and Reource Economic 18: Fiher-Vanden, K., G. H. Jefferon, H. Liu, and Q. Tao What i driving China declined in energy intenity?. Reource and Energy Economic 26: Greene, W. H. and D. A. Henher A latent cla model for dicrete choice analyi: contrat with mixed logit. Tranportation Reearch Part B 37: Henher, D. A., J. M. Roe, and W. H. Greene Applied Choice Analyi: A Primer. Cambridge Univerity Pre. He S., M. Bierlaire, and J. W. Polak Etimation of value of travel-time aving uing mixed logit model. Tranportation Reearch Part A 39: Howarth, R. B, B. M. Haddad, and B. Paton The economic of energy efficiency: inight from voluntary participation program. Energy Policy. 28: Huber, J. and K. Zwerina The importance of utility balance in efficient choice deign. Journal of Marketing Reearch 33: Johnton, R. J., C. R. Weell, H. Donath, and F. Ache Meauring conumer preference for ecolabeled eafood: an international comparion. Journal of Agricultural and Reource Economic 26: Loureiro, M. L., J. J. McClukey, and R. C. Mittelhammer Will conumer pay a premium for eco-labeled apple?. Journal of Conumer Affair 36: Louviere J. J., D. A. Henher, and J. D. Swait Stated Choice Method: Analyi and Application. Cambridge Univerity Pre. McFadden D Conditional logit analyi of qualitative variable in econometric, Zarembka P. (ed.), Frontier in Econometric, New York, Academic Pre, McFadden, D. and K. Train Mixed MNL model for dicrete repone. Journal of Applied Econometric 15: Moon, W., W. J. Florkowki, B. Brückner, and I. Schonhof Willingne to pay for environmental practice: implication for eco-labeling. Land Economic 78: Morey, E., J. Thacher, and W. Breffle Uing angler characteritic and attitudinal data to identify environmental preference clae: A latent-cla model. Environmental and Reource Economic 34: Morri, L. A., M. Hatak, and M. B. Mazi Conumer comprehenion of environmental advertiing and labeling claim. Journal of Conumer Affair 29: Revelt D. and K. Train Incentive for appliance efficiency in a competitive energy environment: random parameter logit model of houehold choice. Review of Economic and Statitic 80: Roe, B., M. F. Teil, A. Levy, and M. Ruell US conumer willingne to pay for green electricity. Energy Policy 29: Ruell, C. S., S. Krarup, and C. D. Clark Environment, information and conumer behaviour: an introduction. in S. Krarup and C. S. Ruell (ed), 19

20 Environment, Information and Conumer Behaviour. Edward Elgar Publihing. Shen, J A Review of Stated Choice Method. International Public Policy Reearch 10(2): Shen, J. and T. Saijo The ocioeconomic determinant of individual environmental concern: Evidence from Shanghai data. Dicuion Paper in Oaka School of International Public Policy (OSIPP), DP-2007-E-003, Oaka Univerity. Stern, P. C Information, incentive, and proenvironmental conumer behavior. Journal of Conumer policy 22: Swait, J A tructural equation model of latent egmentation and product choice for cro-ectional revealed preference choice data.. Journal of Retailing and Conumer Service 1: Swait, J Advanced choice model. in B. J. Kanninen (ed),valuing Environmental Amenitie Uing Stated Choice Studie. Springer. Swait J. D. and Louviere J. J The role of the cale parameter in the etimation and comparion of multinomial logit model. Journal of Marketing Reearch 30: Teil, M. F., B. Roe, and R. L. Hick Can eco-label tune a market? Evidence from dolphin-afe labeling. Journal of Environmental Economic and Management 43: Thøgeren J Pychological determinant of paying attention to eco-label in purchae deciion: model development and multinational validation. Journal of Conumer Policy 23: Tietenberg, T. and D. Wheeler Empowering the community: information trategie for pollution control. in H. Folmer, H. Landi Gabel, S. Gerking and A. Roe (ed), Frontier of Environmental Economic. Edward Elgar Publihing. Train, K Recreation demand model with tate difference over people. Land Economic 74:

21 Figure 1. China Energy Conervation Label Figure 2. China Energy Efficiency Label 21