Preferences for electricity supply attributes in

Energy for Sustainable Development 21 (2014) 89–99

Contents lists available at ScienceDirect

Energy for Sustainable Development

Preferences for electricity supply attributes in emerging megacities — Policy implications from a discrete choice experiment of private households in Hyderabad, India☆ Julian Sagebiel a,⁎, Kai Rommel b,1 a b

Humboldt-Universität zu Berlin, Department of Agricultural Economics, Divison of Economics of Agricultural Cooperatives, 10099 Berlin, Germany International School of Management (ISM) GmbH, Energy Management, 44227 Dortmund, Germany

a r t i c l e

i n f o

Article history: Received 14 November 2013 Revised 9 May 2014 Accepted 5 June 2014 Available online xxxx Keywords: Urban electricity supply Discrete choice experiments Stated preferences Scale-adjusted latent class model India

a b s t r a c t The Indian economy struggles with electricity supply deficits and low quality supply. Although several initiatives including demand side management measures have already been implemented, consumers from different backgrounds suffer from various drawbacks of quality supply. This paper explores the valuation of electricity quality from the perspective of domestic consumers in Hyderabad, India. We conducted a discrete choice experiment with 798 urban households. For analysis, we apply a scale-adjusted latent class model to identify heterogeneity in preferences and in variance-scale. The results confirm the hypothesis of highly heterogeneous household preferences and reveal limited preparedness of domestic users to pay for improved electricity quality and renewable energy. Further, most respondents prefer state owned distribution companies to private enterprises or cooperative societies. We argue that the estimated preferences, implying demand and willingness to pay for single attributes of electricity quality, can help policy makers to adequately incorporate consumers' interests into decision making. The results further indicate that domestic tariff hikes should not be used to finance extension of renewable energies or infrastructure investment to improve reliability in supply. © 2014 International Energy Initiative. Published by Elsevier Inc. All rights reserved.

Introduction It is often argued that, in countries of the Global South, continuous electricity supply is an important prerequisite for economic development and poverty reduction. In reality, however, country-wide electrification in these emerging economies is often not feasible and even within electrified areas, huge differences in received electricity quality remain. In order to overcome the supply gap, governments and policy makers frequently consider investment strategies for electricity infrastructure improvements based on a given budget. Theoretically, the decision maker should distribute each unit of expenditure in a way that it generates the highest marginal benefit. In the electricity sector, observed prices are not adequate to indicate the benefits as electricity infrastructure is not a purely private good and hence no competitive market ☆ Earlier versions of this paper have been presented at the IAEE Conference 2011 in Vilnius, at the IASC Conference 2012 in Hyderabad and at the IAEE Conference 2012 in Stockholm. We would like to thank the audiences for their useful comments. ⁎ Corresponding author at: Humboldt-Universität zu Berlin, Faculty of Life Sciences, Department of Agricultural Economics, Division of Economics of Agricultural Cooperatives, Luisenstrasse 53 D-10099 Berlin, Germany. Tel.: +49 3020936547; fax: +49 3020936501. E-mail addresses: [email protected] (J. Sagebiel), [email protected] (K. Rommel). 1 Tel.: +49 23197513976; fax: +49 23197513939.

exists. While data on electrification and on the quality received by the consumers are widely available there is limited understanding on how this quality is perceived by the consumers. This piece of information can be of critical importance when deliberating policy options and tariff orders. In highly regulated markets, like several electricity markets all over the world, sustainable infrastructure investments and quality improvements cannot be provided efficiently with market-based instruments. It is impossible to observe precisely the preferences of consumers, as they have no or only very limited ways to reveal them. Hence, policy decisions are often based on surveys, secondary data, estimations and projections and, in worse cases, individual opinion and corruption. Often it remains unclear who is most affected by policy changes and how a new regulation effects consumers. Cost–benefit analyses are only possible if there are reliable data on the benefits. Especially for domestic consumers (private households) the benefits from improved electricity quality are rarely observable and affected by various attributes of electricity as final product such as the occurrence of shortages and the share of renewable energies. In this paper we try to contribute filling this gap by applying a discrete choice experiment (DCE) to elicit domestic consumers' preferences for different attributes of electricity quality. The results can be used to adjust power tariffs in the Indian state Andhra Pradesh (AP) in order to reflect individual willingness to pay (WTP) values and to extract additional WTP values for specific attributes related to electricity supply quality. The survey was conducted in February

http://dx.doi.org/10.1016/j.esd.2014.06.002 0973-0826/© 2014 International Energy Initiative. Published by Elsevier Inc. All rights reserved.

90

J. Sagebiel, K. Rommel / Energy for Sustainable Development 21 (2014) 89–99

2010 in the capital of AP, Hyderabad covering 798 households. Attributes were chosen pertaining to the relevance for future policy decisions concerning the quality of electricity supply and sensitive to consumers' utility. The rationale is as follows: If there are distinct preferences and hence high WTP values for certain aspects of improved electricity quality consumers retain disposable income which they could spend for these quality improvements i.e. they would be willing to pay a higher electricity tariff for better electricity supply. Assuming that the total costs for these improvements were lower than the aggregated WTP investments in infrastructure and quality improvements could be financed with higher tariffs without reducing the welfare of consumers. Contrary low WTP values indicate that investments may not be financed by increased tariffs because the costs of these improvements would be larger than the benefits generated by it. For estimation, we apply a scale-adjusted latent class logit (SALC) model which identifies different preference classes within a given sample. This statistical method permits a more exhaustive analysis and takes into account that respondents differ in preferences as well as in the certainty of their decisions. The method seems reasonable as the sample consists of several subgroups including different income groups, religions, educational backgrounds, etc. The results suggest, firstly, that consumers are highly heterogeneous. Secondly, more than 85% of the respondents are not willing to accept further tariff hikes even if power cuts reduce and/or the share of renewable energy increases. Thirdly, many respondents are satisfied with governmentowned distribution companies, while about 10% strongly favor private or co-operative distribution companies. The paper is structured as follows: In Section 2 we introduce the power sector in AP. Section 3 explains the DCE and the SALC model. Section 4 overviews the data and the survey details. In Sections 5 and 6 we illustrate the results, discuss the WTP values and interpret the class characteristics. Section 7 gives a short analysis of the interaction of socio-economic variables with the preference classes and Section 8 concludes with recommendations for improvements of the efficiency of the electricity tariffs in AP.

aims to solve the various deficits of the power sector such as growing demand surplus, increasing CO2 emissions, low share of renewable energy for power generation, power supply interruptions, and low energy efficiency and overuse (CESS, 2013). In the capital Hyderabad, excess demand is still growing, burdening the local energy infrastructure, and leading to unscheduled power cuts and large voltage fluctuations (Sreekumar et al., 2007). The rapid growth of Hyderabad's population restricts the development of the infrastructure. The number of household connections to the distribution grid increases with an annual rate between 8 and 9% and the rising industrialization of the urban areas again contributes to the rapid growth of the demand surplus in the electricity market. Domestic end use tariffs, i.e. the price per KWh of electricity consumed by domestic households, are determined by the Andhra Pradesh Electricity Regulatory Commission (APERC), and increase with the total consumption per connection. Consumers using less than 50 KWh per month pay 1.45 INR per KWh, while those who consume more than 500 KWh per month pay 8.38 INR per KWh. This tariff structure should relief financial burden from the low income classes but has been heavily criticized that it does not (Reddy and Raghu, 2012). We argue that the current tariff system does not produce efficient outcomes. If tariffs do not consider all costs and benefits of power generation, transmission and distribution, they fail to set sufficient incentives for efficiency investments of both commercial and domestic consumers. The construction of tariff structures considering these features requires knowledge of consumer preferences for all attributes of electricity utilization. The problem of market observation is that revealed preferences, based on regulated prices, do not reflect the complete set of consumer preferences. An optimal tariff is reached if it reflects all components of the utility functions of consumers. This includes also the source of generation and the organizational form of the distribution company (Sagebiel et al. 2014). Discrete choice experiments Background of discrete choice experiments

Overview of the power sector in Andhra Pradesh Electricity demand in AP as well as in India has been growing continuously faster than generation capacities over the last decades. The Indian economy suffers from permanent power cuts and insufficient energy infrastructure (Tongia, 2007; Lal, 2006). In 2001–2002, the peak deficit in AP was 19.9%. It dropped to 2.3% in 2004–2005 (Central Electricity Authority, 2011), which was achieved by the introduction of demand side management measures, limited supply for agriculture and a stricter control of the distribution companies (Deb et al., 2012). However, due to a strong increase in demand it raised steadily to 20.2% in 2012–2013 (Central Electricity Authority, 2013), leading to increased scheduled and unscheduled power cuts. The current share of coal fired power plants in AP is about 50% i.e. 8783 MW (Central Electricity Authority, 2014c). In 2012–2013, the total net generation from thermal power plants summed up to 117,231 GWh with a total number of CO2 emissions of 92.14 million tons, reflecting a weighted average emission factor of 0.785.2 Overall India, new capacities come mainly from fossil fuels, accounting for comparatively high CO2 emissions. In 2014 additional thermal power projects with a total capacity of 15,234 MW are planned (Central Electricity Authority, 2013). In AP, limited financial capabilities and governance failures impede public and private investments in energy efficient technologies and renewable energies. The 12th Five Year Plan of AP (2012–2017) 2 The Central Electricity Authority (CEA) in India reports four different types of emission factors (Central Electricity Authority, 2014b). The weighted average emission factor is calculated as kilograms of CO2 emitted per megawatt-hour produced over all power plants. We used the data provided by CEA (Central Electricity Authority, 2014a) to calculate weighted average emission factor for AP.

The DCE method is a survey based instrument to elicit preferences, choice probabilities and WTP values for characteristics or attributes of a good. Respondents are repeatedly asked to choose between alternatives which include these attributes with associated attribute levels. The attribute levels vary over the alternatives. A respondent usually answers six to 16 choice sets and the number of attributes rarely exceeds eight. Fig. 1 depicts an example for a choice set card which has been used in this study. The selection of the attributes and levels is challenging. If attributes are irrelevant to the respondent or dominated by other attributes or if levels are too close or too far away from each other, the external validity and the estimation are at risk. Usually extensive pretesting and focus group discussions before the experiment are conducted to optimally design the choice sets. DCEs can be carried out online, per post or with inhouse interviews. After collecting the data several econometric models are available for estimation. The underlying economic theory goes back to the contributions of Lancaster (1966) and Rosen (1974) to consumer theory. Thurstone (1927) laid the foundation for the random utility model and Manski (1977) formalized it as the theoretic basis for the econometric modeling. The most frequently applied econometric model is the conditional logit (CL) model (McFadden, 1974) but its use is restricted by several strong assumptions. A more flexible formulation is the random parameters logit (RPL) model (e.g. Revelt and Train, 1998; Hensher and Greene, 2003) which assumes the parameters to vary randomly across individuals. The RPL captures heterogeneity in preferences and allows calculating individual parameters. Semi-parametric variants of the RPL are the latent class logit (LC) models (e.g. McCutcheon, 1987; Greene and Hensher, 2003). Here, heterogeneity is assumed to be discrete and limited to a number of classes. In the current debate in the


91

Choice Set 1 No

Alternave 1

Alternave 2

Duraon of scheduled

Summer: 15 minutes/day


power cuts

Winter: 5 minutes/day


Duraon of unscheduled



power cuts



5 % renewable

10 % renewable

Government (APCPDCL)

Private

0 % increase

10 % increase

1

2

Renewable energy in 4 energy mix 5

Instuonal set up Addional costs per

6 month Please ck one opon

Fig. 1. Example of a choice set card.

literature various modifications to incorporate heterogeneity in discrete choice models are discussed. These include mixing distribution techniques (Fosgerau and Hess, 2009; Campbell et al., 2010), combinations of PRL and LC models (Bujosa et al., 2010), and more elaborated approaches to include attitudes in the estimation process (Hess and Beharry-Borg, 2012). In this paper we apply an extension of the LC model, the SALC model. It will be derived in the following section. The scale adjusted latent class logit model The SALC model can be interpreted as a generalization of the CL model and is the semi-parametric version of the generalized multinomial logit model.3 In the simple LC model, an additional set of parameters is estimated with a multinomial logit model to differentiate a given number of groups or classes of individuals. Assume, for example, that the sample consists of two groups which differ in preferences towards a policy intervention. One group strongly favors the intervention and the other one is strongly against it. The LC model classifies the respondents into these two groups based on their choices. Each individual is assigned with a probability to be member of one class, thus preference heterogeneity is captured discretely in these two classes. However, the model does not consider that individuals may also vary in their ‘certainty’ of decisions. While some respondents are very clear on what they are choosing (for example because they are well aware of the topic) others choose more randomly among the alternatives.4 That is, the error term has a high variance (and hence a small variance-scale parameter). Neglecting this possibility could lead to a situation where respondents 3 The generalized multinomial logit model is an extension of the RPL model (Fiebig et al., 2010). It assumes parametric distributions over individuals of the preference and relative variance-scale parameters. However, as discussed in Hess and Rose (2012), the validity of such models is contested. 4 This is an interpretation that is not totally in line with the underlying random utility model. In the random utility model, we assume that the choice is fully deterministic, but the researcher cannot observe everything. Hence, respondents with a high randomness in their decisions should be interpreted as respondents with a large unobserved part of utility. We, however, stick to the above interpretation as it is more useful in this context.

have identical preferences but are assigned to different classes due to differences in variance-scale. As the variance-scale and preference parameters are confounded in discrete choice models, the LC model is not able to distinguish between an individual with a large preference parameter and small variance-scale and an individual with a small preference parameter and a large variance-scale. Magidson and Vermunt (2008) developed the SALC model which estimates the relative variance-scale and the preference classes separately. Earlier, the model has been applied for example to health economics (Flynn et al., 2010) and tourism analysis (Burke et al., 2010), yet we are not aware of an application in low-income countries and in the energy sector. The SALC model can be formally described as follows: Assume a randomly selected individual i who chooses between n alternatives in t choice situations. Each alternative accommodates k attributes with Aiknt levels which vary over alternatives. Each alternative creates a certain amount of utility and the respondent chooses the alternative which corresponds to the highest level of utility. The utility of each alternative is quantified in indirect utility functions.5 For simplicity we assume utility functions Uint for each alternative n and individual i and choice situation t to be linear with respect to attribute levels Aiknt. The utility function can be separated into two parts. There are utility sensitive elements that can be directly observed, i.e. effects of the attributes on utility. These are described in Vint and represent the deterministic part of the utility function Uint. The elements that are not observed by the researcher are comprised in the term eint. eint is the error term and varies randomly. It comprises all effects that influenced the decision but cannot be attributed to the explanatory variables. Examples would be the unobserved attitudes of the respondent or his current mood that influences his decision. This formulation can be written as U int ¼ V int þ eint ¼ λβ1 Ai1nt þ λβ2 Ai2nt þ … þ λβk Aiknt þ eint

ð1Þ

5 In microeconomics, one distinguishes between direct and indirect utility functions. A direct utility function describes the utility which stems from the consumption of a bundle of goods. An indirect utility function is the result of an optimization process. It describes the highest level of utility achievable under a given income and prices. In discrete choice modeling the indirect utility function is usually used (Alpizar et al., 2001).

92


where the β1 to βk are the corresponding utility parameters. The variance of eint is normalized to VAR(eint) = π2/6 by multiplying with scale λ. In the CL model, it is assumed that the eint is independent and identically distributed over all individuals and alternatives. This implies that the scale λ has the same magnitude for each i and t. The CL probability for individual i to choose alternative m in choice situation t is given as expðλβ 1 Ai1mt þ λβ2 Ai2mt þ … þ λβk Aikmt Þ Pr imt ¼ XN expðλβ1 Ai1nt þ λβ2 Ai2nt þ … þ λβk Aiknt Þ: n¼1

ð2Þ

It is not possible to distinguish between λ and β. In the estimation, β⁎ = λβ is obtained. The SALC model can be regarded as a generalization of the CL model with βk taking a finite number R of values 〈βk|1, βk|2, …, βk|R〉 with corresponding probabilities 〈h1, h2, …, hR〉 for different preference classes and a finite number S of values 〈λ1, λ2, …, λS〉 with corresponding probabilities 〈g1, g2, …, gs〉 for the variance-scale classes. The unconditional probability to choose alternative m is the weighted average of the r βk|r and s λs parameters. Pr imt ¼

S X R X

g s hr Pr imtjsr

ð3Þ

s¼1 r¼1

with Primt|sr being the CL probability to choose alternative m when belonging to variance-scale class s and preference class r.

Pr imtjsr

exp λs βi1jr Ai1mt þ λs βi2jr Ai2mt þ … þ λs βikjr Aikmt ¼ XN exp λ β A þ λ β A þ … þ λ β A s i1jr i1nt s i2jr i2nt s ikjr iknt n¼1

ð4Þ

gs and hr are unknown but can be estimated with a multinomial logit model. expðζ s X i Þ g s ¼ XS expðζ s X i Þ s¼1

ð5Þ

and exp ηr Z i hr ¼ XR exp ηr Z i r¼1

ð6Þ

where Xi and Zi are the vectors of case-specific variables like income, age, or attitudes that have an effect on the variance-scale and preference probability and ζs and ηr are the corresponding parameter vectors for variance-scale class s and preference class r. The vectors Xi and Zi can comprise only a constant if case-specific variables are not available or do not explain the probabilities. The number of classes can be chosen by the researcher based on his assumptions, a-priori findings and goodness-of-fit statistics. The reason for the choice of this model came from practical considerations. First, discrete heterogeneity is easier to understand for a layman in choice modeling. Especially when it comes to informing policy makers, a clear cut separation into classes is more convincing. Second, including a scale parameter reduces the confounding of preference classes. In a situation where many respondents faced difficulties with the complexity of the choice sets, not accounting for variance-scale can lead to severe misclassification. In this study, there is reason to assume differences in the scale parameter as many respondents were uneducated and illiterate. They faced more difficulties in choosing and their choices were more prone to random decision making. Third, compared to the models described above, the estimation process is relatively simple and less prone to defective results.

Discrete choice experiments in the energy sector While there is a large amount of literature on WTP for better and renewable energy, limited research has been undertaken to investigate preferences for better electricity quality in terms of reliability in service. In the following we will present some studies that focused on these aspects of quality thereby neglect the large amount of literature on renewable energy in developed countries. Hanisch et al. (2010) investigated strategies of power customers in emerging megacities coping with power scarcity. They defined the status of consumers on power markets in Hyderabad as a social dilemma and examined relevant drivers of consumer decisions to cope with power scarcity. The majority of the 142 sampled private households preferred quality improvements of power supply over increases in delivered quantity. Abdullah and Mariel (2010) investigated consumer preferences for improvements of power supply quality in Kisumu, Kenya. Their paper focused on the specific situation of electricity markets which are characterized by monopolistic market structures and frequent power cuts during the day. Using CL and RPL models on DCE data from a survey of electrified rural households, they investigated unobserved and observed heterogeneity with interaction terms between socio-economic characteristics, and costs. The choice set included the attributes number and length of planned power cuts, type of distribution provider, and costs of electricity consumption. For the attributes frequency and durations of power cuts the results of the 808 respondents indicated preference heterogeneity among the households, i.e. preferences for better quality varied among the respondents, which could be explained with socio-economic variables. Amador et al. (2013) studied WTP values for different electricity suppliers in the Canary Islands. They investigated reliability of service, share of renewable energies and availability of energy audit services. Using a RPL model with error components, they find significant WTP values for the reduction of power cuts especially from persons who have past experiences with power cuts. The WTP for renewable energies increases with knowledge on greenhouse gas emissions and energy saving behavior. The aggregated WTP values for renewable energies are about 50% of the costs to install them. They conclude that market mechanisms would not suffice to increase the share of renewable energies. Carlsson and Martinsson (2008) investigated preferences of Swedish consumers for reductions in power cuts. They used a RPL model to account for unobserved heterogeneity. The DCE on power cuts was embedded in a questionnaire on power consumption and socio-economic characteristics. 473 respondents stated preferences for reductions of power cuts. In the study, power cuts were differentiated between summer and winter time and between weekdays and weekend. The results indicate higher WTP for avoiding a power cut during weekends in winter and increasing values with the duration of the power cut. Our study is unique in the sense that we create a direct link of different packages of electricity quality to prices. While most studies focus on specific attributes in more detail, we take into account several, unrelated attributes that describe quality for the consumers we are studying. This comes at the cost at less specific results for each attribute, yet provides a wider overview of relevant aspects.

Survey and model specification Sample and questionnaire description The survey had been conducted in February 2010 in Hyderabad as part of the research project ‘Sustainable Hyderabad’.6 It covered 798 domestic electricity consumers within the limits of the Greater Hyderabad Municipal Corporation area. The sample had been stratified by income groups. We developed these groups based on individual data from the local distribution company and a previous study conducted in 2009 6

More information on the project is available on www.sustainable-hyderabad.de.


(Hanisch et al., 2010). The latter one linked income groups with electricity consumption and the former one provided the distribution of electricity consumption within the study area. As a result, 10% high class, 50% middle class and 40% lower class households were targeted as this corresponds to the distribution of all registered domestic consumers in Hyderabad. For three days, a group of 16 enumerators were trained by the authors and a local consultant. The pretests consisted of 100 interviews and confirmed a clear understanding of the structure of the DCE and of the single attributes. After refinement of the questionnaire, the enumerators had been sent to different locations and interviewed respondents in accordance to the stratification.7 The questionnaire was split into three parts. The first part investigated the energy consumption patterns. The second part included the DCE and DCE relevant questions and the third part covered questions regarding attitudes towards reformation of the energy sector, renewable energy and climate change.

Development of attributes To identify the relevant attributes, we relied on four prerequisites. Firstly, we exploited an explorative energy consumption study from Hyderabad (Hanisch et al., 2010). The study investigated problems of private households related to electricity supply, WTP for reduced power cuts, and the status quo supply situation. Secondly, we conducted a small survey with 30 representatives of different areas in Hyderabad asking for a ranking of problems and the current status of their electricity supply. Thirdly, we performed pretests with different combinations of attributes and asked the respondents in focus group discussions about their opinion on our choice of attributes. Fourthly, experts in the electricity sector were interviewed concerning their opinion on different attributes and levels, after presenting the questionnaire to them. A thorough analysis of the results and further pretests led to the following attributes and corresponding levels in Table 1. We selected the attributes in order to reflect the four main criteria of electricity supply concerning consumer preferences. The first criterion is availability of electricity, hence we used the duration of power cuts per day as an indicator. We varied the levels of these only in summer because it turned out that the biggest problem with power cuts is the non-availability of cooling systems. Deciding on the levels for power cuts was problematic. The official data from the Central Electricity Authority, India (Central Electricity Authority, 2009) contradicted the results from the explorative study and the focus group discussions. The former stated an average duration of power cuts of 10 min per day while the latter perceived power cuts between 60 and 120 min per day. Hence, 60 min was chosen as the worst case level (30 min scheduled plus 30 min unscheduled) and no power cuts as the optimal level. The pretests confirmed that unscheduled power cuts had been perceived differently by the consumers. Thus, we distinguished between scheduled and unscheduled power cuts. The second criteron of electricity supply reflects – independent of the first criterion – external effects of power generation. External effects, in this context, refer to the CO2 emissions caused by the generation of electricity. Under the assumptions of neoclassical economics, these adverse effects are not taken into account by electricity consumers. In this study, we wanted to investigate in how far consumers are willing to pay to reduce the external effects. We reduced the complexity of

7 The clustered and stratified sample has the major drawback that it is not fully random, as addresses or registers of the population were not available. Hence, it was impossible to preselect the sample. Instead, enumerators were instructed to randomly select households based on predefined criteria. This ‘enumerator bias’ could, for example, lead to a negligence of unfavorable households that are not easily accessible, seem unfriendly or far away. In case these neglected households differ strongly in respective characteristics, the sample is biased.

93

external effects of power generation by using the share of renewable energy in the electricity mix. The current share of renewable energy in Hyderabad is 2%, however APERC sets the minimum requirement to 5% (Andhra Pradesh Electricity Regulatory Commission, 2012). Our expert interviews revealed that in the near future, a maximum of 10% could be achieved.8 We considered these three options as realistic scenarios and incorporated them into the choice experiment. The third criterion which we analyzed in our DCE is the organization of the power sector. Mueller and Rommel (2011) discuss the implications on different governance structures in the electricity sector including the advantages of cooperatives over investor owned firms and Sagebiel et al. (2014) find that the organizational form partly determines the consumers' choice of distribution companies. Hence, the organizational form of the distribution company might play a crucial role for the customer in our context as well. Expert interviews and a survey with representatives from different communities indicated that consumers may expect that supply quality and quantity are determined by the organizational form, in other words, private or cooperative distribution companies may produce different standards compared to government-owned distribution companies. Incorporating the organizational form of the distribution company sheds light on the preferences for reform and market liberalization. The status quo, a governmentowned distribution company, can be substituted with either private, profit maximizing companies or with co-operative societies, where the consumers are members and can participate in the decision making process. The costs for improved electricity supply quality build the fourth criterion of electricity supply in our analysis. Reductions in power cuts and investments in renewable energy lead to additional costs. The WTP for the preferred attributes indicates the additional costs for these attributes, which can be covered by the market price if WTP values are absorbed completely. The additional costs per month are given in percentages and derived from expert interviews and the explorative study, where consumers were asked to state their WTP for improvements in electricity quality. The amount never exceeded 20% of the electricity bill. Having created the attributes and its levels, there are totally 35 = 243 alternatives. We then produced an orthogonal array with 54 alternatives and randomly generated 27 choice sets with two alternatives per choice set. We blocked the treatment into three parts with nine choice sets pre survey in order not to tax the respondents' patience.

Results Inspection of data In the following, we describe selected socio-demographic variables that were collected in the survey (Table 2). AGE and AGESQ are the age and the squared age of the respondent, respectively. The average age is 34.26 years and the standard deviation is 10.5 years. HHMEMBERS refers to the number of household members in the respondent's household. It ranges from one to 17 members with an average household size of five members. CATEGORY describes the social status of the household which is divided into high class (1), middle class (2) and slum (3). As the sampling was based on the category, the proportions of CATEGORY are similar to the results from secondary studies. 9.19% come from high class, 42.7% from middle class and 48.11% from slums. SEX refers to the sex of the respondent. Females are coded with 0 and men with 1. The sample is relatively even distributed

8 AP is part of a regional grid, which has a fixed share of renewable energies. However, the distribution companies in AP buy electricity from different sources and exhibit some freedom in their choices. APERC determined that each distribution company has to buy 5% from renewable sources, but, in the current situation, the distribution companies do not follow this instruction completely. This is the reason why Hyderabad has only a share of 2%.

94


Table 1 Attribute description. Variable

Description

Levels

Code

SCHED

Scheduled power cuts per day in minutes

UNSCH

Unscheduled power cuts per day in minutes

REN

Percentage of renewable energy in the electricity mix

PRIV

Whether supply is carried out by private company

COOP

Whether supply is carried out by co-operative society

COST 10%

10% additional costs to monthly electricity bill

COST 20%

20% additional costs to monthly electricity bill

30 min 15 min 0 min 30 min 15 min 0 min 2% 5% 10% Private company COOP or government Co-operative PRIV or government 10% No add. costs or 20% 20% No add. costs or 10%

0 1 2 0 1 2 0 1 2 1 0 1 0 1 0 1 0

with 52.27 percent female respondents. BACKUP indicates whether the respondent has a diesel generator or inverter as a backup system to overcome power cuts (1) or not (0). 13% of the respondents use a backup system. INCPERHEAD is the income per head for the household divided by 1000. The average income per head is 4360 INR, with a minimum of 200 INR and a maximum of 200,000 INR. The standard deviation is 7606 INR. ELECTRICITYEXP is the average monthly expenditure for electricity divided by 100. The average expenditure is 379 INR with a minimum of 42 INR and a maximum of 3042 INR and a standard deviation of 311 INR. Although several other variables have been collected the analysis will be limited to the above described variables due to space limitations. Before conducting the regression analysis, it is useful to inspect the DCE data and identify irrational behavior. Further, instead of assuming the decision process being led by the linear utility function described above, one could analyze in how far respondents followed other decision rules. We define a decision rule as a heuristic which helps the respondent to choose an alternative. An example for a decision rule is: Always choose the alternative with the lowest price. We analyze the simplest case of persistently choosing the alternative where one specific attribute is better than in the other alternatives. This behavior is also known as having lexicographic preferences. Table 3 summarizes these decision rules. The first column states the decision rule, the second and the third column count the respondents who applied the rule in absolute numbers and percentages. The next columns provide the average value of three key variables AGE, BACKUP and INCPERHEAD of those who used the decision rule. 257 respondents (32.37%) decided by the price only i.e. the cheaper alternative has always been chosen, regardless of all other attribute levels. This is a clear indication – and it will be confirmed later – that a large share of the population is not willing to pay for

Table 2 Socio-demographic variables. Variable

Mean

Std. dev.

Min.

Max.

N

AGE AGESQ HHMEMBERS CATEGORY SEX BACKUP INCPERHEAD LNINCPERHEAD ELECTRICITYEXP LNELECTRICITYEXP

34.26 1283.83 4.89 2.39 0.48 0.13 4.36 0.98 3.79 1.09

10.49 817.97 1.77 0.65 0.5 0.34 7.61 0.85 3.11 0.67

18 324 1 1 0 0 0.2 −1.61 0.42 −0.87

73 5329 17 3 1 1 100 4.61 30.42 3.42

762 762 794 794 794 794 722 722 778 778

improved electricity quality and for an increased share of renewable energy.9 They do not seem to differ in AGE compared to the overall sample, but have lower average income per head (3680 INR compared to 4360 INR). Only 8% have a backup system, which again is below the sample average of 13%. The preferences for the distribution company seem also to determine behavior. About 11% of the respondents have always chosen the alternative with a governmentowned distribution company, i.e. the current situation. These respondents have a higher average income per head of 5580 INR and only 8% have a backup system. About 7% acted oppositely, choosing always a private or co-operative distribution company. Respondents who favor a private supplier have a relatively high average income per head of 5790 INR, while those who prefer a co-operative supplier have a relatively low average income per head of 2550 INR and are relatively young on average. Decision rules applied to the other attributes are negligible, accounting for less than 5% of the sample. Apart from these heuristics, we tested for counterintuitive or ‘weird’ preferences. We found that 15 respondents (1.89%) always decided opposite to the heuristics defined earlier. For example, some respondents always picked the alternative with the higher price. This behavior, however, is not necessarily inconsistent. Firstly, a respondent could (falsely) assume that a higher price implies better quality and hence chooses always the higher price option. Secondly, the design of the choice sets and the blocks may accidentally lead to a solution where the seemingly irrational behavior emerges although it is fully consistent with utility maximizing behavior. Thirdly, it could be the case that these preferences are the true preferences. For example, the respondent could be afraid that the tariff will increase in the long run if the share of renewable energy increased. This assumption might let him choose the alternative with less renewable energy. We incorporated another method into the experimental design to identify inconsistent choices: In one specific choice set, the first alternative dominated the second one, i.e. for all attributes, the first alternative was better except for organizational form which is a nominal attribute. With this test, we identified 46 respondents who can be categorized as ‘weird preferences’.10 They have an average income of 3980 INR and 9% use a backup system.

9 An expert interview with the deputy director from the Andhra Pradesh Electricity Regulatory Commission revealed that no customer in whole AP has ever inquired the existing green tariff. We found that most people are not aware that this tariff exists. Thus, it could be that the lack of demand is due to an information deficit. 10 Some respondents revealed both counterintuitive behavior and inconsistent choices at the same time.

J. Sagebiel, K. Rommel / Energy for Sustainable Development 21 (2014) 89–99 Table 3 Heuristics and weird preferences.

95

Table 4 Estimation results: conditional logit models.

Heuristic

N

%

AGE

BACKUP

INCPERHEAD

Variable

Scheduled power cuts Unscheduled power cuts Total power cuts Renewable energy Additional costs COOP supplier Private supplier Government supplier Sum of heuristics Counterintuitive heuristics Inconsistent behavior Sum of weird preferences

17 4 8 6 257 15 31 76 414 15 34 46

2.14 0.5 1.01 0.78 32.37 2.36 4.87 11.95 52.14 1.89 4.28 6.17

35 29 29 33 34 31 36 34 33 38 33 35

0.18 0.25 0.13 0.17 0.08 0.13 0.13 0.08 0.1 0.07 0.09 0.09

3.25 2.43 9.83 3.46 3.68 2.55 5.79 5.58 4.19 4.01 4.46 3.98

SCHED

Bold letters indicate the sum of heuristics and the sum of weird preferences.

Regression analysis The descriptive analysis above provides a first overview of the data. This is important because the results of the regression models can be verified. If regression analysis gives counterintuitive results, we might have done something wrong or the model is misspecified. For example, it might be the case that some respondents have opposite preferences to other respondents. The parameter to be maximized is hence the mean of these preferences if, for example, the CL is used. This results in wrong estimators, even though the estimation seems statistically satisfactory. Hence, it is often helpful to ‘assist’ the maximization process by inducing prior knowledge in the model. In a first step, we estimate the CL model and then develop the SALC specification that captures the implied heterogeneity. Preference classes are endogenously developed to detect different structures of utility. Table 4 gives the CL results for all respondents (Model 1), for all respondents excluding the respondents with weird preferences (Model 2) and for all lexicographic respondents (Model 3). The structure of the utility function needs further attention. As of now, we assumed linear effects of the attributes on utility, i.e. a change from level 1 to level 2 has the same effect as from level 2 to level 3. Intuitively, this is not always useful. For example might an increase in costs by 10% be regarded as acceptable, and hence does not have an effect on utility while an increase to 20% be large enough, so that utility would decrease strongly. The usual strategy to account for non-linearities is to incorporate dummy variables. When using dummy variables, each level is assigned to a separate variable that is one if the level is present and zero otherwise. By doing so, the effect for each level is estimated separately. If the estimated parameters of the dummy variables differ, non-linearity can be assumed. We tested for non-linear effects of the continuous variables on utility by introducing dummy variables. We found that COST is strongly non-linear in utility. For SCHED, UNSCHED and REN, we did not find evidence for a nonlinear relationship.11 Hence, COST was dummy coded, all other numerical attributes were treated as continuous variables. We recoded SCHED and UNSCHED from 0 15 30 to 2 1 0, respectively to increase visualization of estimates.12 The recoding of REN needs further explanation. As described above, the three levels are based on the status quo, the governmental minimum requirement, and the maximum that is realistic in the near future. In this manner, the attribute was explained to the

11 We used Wald tests to investigate non-linear effects. A Wald test tests whether several parameters are equal and is frequently used in models that are estimated with the maximum likelihood method. The interested reader is referred to Hensher et al. (2005). 12 That is, the original coding indicating minutes will lead to parameters based on the effect on a one minute increase. This effect is very small and such coding generates parameters like 0.000…. In the recoding, the estimated effect reflects a 15 minute change, leading to larger parameters.

UNSCH REN PRIV COOP COST 0.1 COST 0.2 N Log likelihood (full) Log likelihood (null) Pseudo R2

Model 1

Model 2

Model 3

0.286⁎⁎ (0.024) 0.039 (0.024) 0.022 (0.021) −0.167⁎⁎

0.378⁎⁎ (0.026) 0.074⁎⁎ (0.026) 0.040† (0.072) −0.131⁎

0.293⁎⁎ (0.039) 0.010 (0.041) −0.042 (0.034) −0.610⁎⁎

(0.051) −0.223⁎⁎ (0.048) −0.519⁎⁎ (0.046) −1.465⁎⁎ (0.045) 7146 −4213.23 −4953.23 14.94%

(0.014) −0.220⁎⁎ (0.051) −0.627⁎⁎ (0.049) −1.632⁎⁎ (0.049) 6732 −3844.23 −4666.37 17.62%

(0.093) −0.654⁎⁎ (0.094) −0.921⁎⁎ (0.086) −2.391⁎⁎ (0.084) 3726 −1788.49 −2582.66 30.75%

Model 1 describes the model with all respondents, Model 2 describes the model excluding weird preferences and Model 3 describes the model only with respondents who followed lexicographic preferences. Log likelihood (full) refers to the log likelihood value at convergence of the respective model. Log likelihood (null) refers to the log likelihood value of the Null model, i.e. a model where the constant is the only parameter estimated. The increase in the Log likelihood value indicates the explanatory power of the estimated parameters. Standard errors in parentheses. Significance levels: †: 10% *: 5% **: 1%.

respondents. During conduction of the survey, we found that respondents perceived the levels on an ordinal scale. This finding was supported by the estimation results. The (maybe unexpected) rescaling leads to a better fit. However, introducing dummies for REN did not improve the likelihood function and a Wald test could not reject the hypothesis of equal slopes. Thus, we recoded 2%, 5% and 10% with 0, 1 and 2 respectively. In all models, the parameters display the expected signs and UNSCH and REN are not significant on a 10% significance level in the full model and in the lexicographic model. The large magnitude of the cost parameter reflects the presence of the heuristics described above. The model fit is slightly improved when we exclude the respondents with weird preferences and the model fit improves strongly when we include only respondents with lexicographic preferences. Both results are not surprising as we deselect respondents who ‘do not fit’ into the ‘mainstream’ preferences.13 Considering the lexicographic model, another drawback of CL model becomes obvious. Assuming that the lexicographic respondents describe the whole population, the CL specification gives us a good idea of the aggregated preference structure in the population. Yet it does not cover the obvious fact, that some respondents follow one heuristic, while others follow another, very different heuristic. In the following, we apply the SALC model described above. This panel data model captures heterogeneity in variance-scale and in preferences. The classes reported by the model are endogenously derived and reflect, similar to cluster analyses, similarities in the decision behavior. The classes in the SALC model should not be confused with classifications derived from other exogenous variables like income groups, religion etc. As mentioned before, it is not straightforward to decide on the number of variance-scale and preference classes but we can use our findings from the descriptive analysis to ‘inform’ the model on what we already know. To do so, we introduce another parameter τr, the ‘known class indicator’, to identify known preference classes. This parameter is zero if a respondent cannot belong to a certain class and 13 It should be mentioned that the results are statistically not valid due to selection bias. However, it demonstrates how the descriptive data analysis can be used to better understand the data.

96


one otherwise. The unconditional probability to choose an alternative then becomes Pr im ¼

S X R X

gs τ r hr Pr imjsr :

ð7Þ

s¼1 r¼1

To select a SALC model specification, several criteria can be considered. In the empirical literature, goodness-of-fit measures like Bayesian Information Criterion (BIC) or Corrected Akaike Information Criterion (CAIC) are regularly consulted as an indication for a good model. Additionally, the usefulness of the estimated parameters related to previous findings from inspection of the data can be considered. Complex likelihood functions, maximizing many parameters simultaneously, are unavoidable in models like the SALC model. They are usually not convex and have several local maxima. Hence, it is possible that an incorrect local maximum, leading to wrong parameter estimates, is identified. The knowledge generated from the inspection of the data gives an indication, whether the model might be wrongly estimated. In this paper, we present the results of a four preference class and two variance-scale class model. The model shows meaningful results which are in line with our previous findings and goodness-of-fit statistics were reasonable compared to other specifications.14 We used the known class indicator for the cost heuristics in class 1 and for the weird preferences in class 3. In detail this means that, a priori, 32.4% (257 respondents) were classified in class 1 and 6.2% (46 respondents) in class 3. All remaining respondents were assigned a probability for all four classes. This means that classes 1 and 3 are not restricted to the respondents with known class indicator as the other respondents can also be classified into these classes. In order not to over complicate the analysis, we neglect the inclusion of socio-demographic data in the model and will discuss it in Section 7. The estimation was carried out with the statistical software package Latent Gold Choice 4.5 with the additional syntax module (Vermunt and Magidson, 2005). Table 5 summarizes the estimated parameters and standard errors. The pseudo R2 comes to 44.81% and the overall model is highly significant. The multinomial logit formula for the classes includes a constant for each class, which is normalized to one for class 1. Class 3 parameters have small magnitudes and are not significant and the class constant is significant on a five percent level. The parameters in class 4 are significant at a five percent level and class 1 and class 2 have highly significant parameter values. The scale classes are relatively even distributed, with the second class having a scale parameter of 3.665 relative to scale class 1. The parameter values have to be multiplied by this factor to determine the parameter value for scale class 2. Members of scale class 2 have a smaller error variance and are hence more deterministic in their choices. The probabilities to be member in classes 1, 2, 3, and 4 are 50%, 5%, 40%, and 4%, respectively. Discussion A closer look into the four preference classes indicates the structure of heterogeneity. Additional to the 32.4% predefined cost-sensitive respondents in class 1, another 19% were assigned to this class, which makes it the biggest one. This class is, as expected, characterized by a relatively high cost parameter and hence displays low WTP values for the attributes. A member of this class would not want a deviation from the status quo if it was related to tariff hikes. Leaving costs aside, the preferences tend towards a reduction of power cuts and an increase in renewable energy. Additionally, a privatization of the energy sector is not desired. We call this class the cost-sensitive conservatives or

14 We tested several models with different known class indicators and classes. Some models showed better goodness-of-fit values than the selected one but had drawbacks otherwise, e.g. suspicious parameter values. These results are available on request from the authors.

Table 5 Estimation results: Latent Class Model. Variable

Class 1

Class size SCHED

0.51 0.524⁎⁎

UNSCH REN PRIV COOP COST (0.1) COST (0.2) Class

Class 2 0.05 4.658⁎⁎

(0.082) 0.218⁎⁎ (0.067) 0.283⁎⁎ (0.058) −0.273⁎⁎

(1.215) 1.384⁎⁎ (0.473) 3.405⁎⁎ (0.876) −4.960⁎⁎

(0.087) −0.371⁎⁎ (0.094) −1.684⁎⁎ (0.272) −3.381⁎⁎

(1.421) −8.9563⁎⁎ (2.423) 1.481† (0.766) −6.464⁎⁎

(0.502) 1 (–)

(1.753) −2.340⁎⁎ (0.191)

Class size Scale factor

Class 3 0.40 −0.008 (0.015) 0.007 (0.014) −0.024† (0.014) −0.025 (0.036) −0.046 (0.029) −0.004 (0.032) −0.056 (0.039) −0.2351⁎ (0.110)

Class 4 0.04 8.940⁎ (3.703) 4.0751⁎ (1.167) −1.102⁎ (0.459) 5.590⁎ (2.43) 5.385⁎ (2.44) −3.904⁎ (1.857) −7.160⁎ (3.113) −2.615⁎⁎ (0.215)

Scale class 1

Scale class 2

0.41 1 (–)

0.59 3.665⁎⁎ (1.034)

N Number parameters Log-likelihood Chi squared Pseudo R2

7146 33 −3738.549 −37086.147 44.81%

Standard errors in parentheses. Significance levels: †: 10% *: 5% **: 1%.

CostSensCons. The second largest class is class 3 with 40%. This class included the predefined weird preferences group. It seems that these respondents have no clear preferences. This is affirmed by the small magnitudes of the parameters in this class. We call this class the indecisive laymen or IndecLaym. Class 2 and class 4 are small and indicate clear-cut preferences. Conspicuous in class 2 is the cost parameter. An increase of the electricity bill by 10% does not play a significant role. However, with an increase to 20%, an unknown threshold is exceeded and costs become important. These members would agree to a tariff hike only if it is less than 20%. Also, we observe strong preferences against privatization of the sector and for an increase of the share of renewable energy and reduction of power cuts. We name members of class 2 green conservatives or GreenCons. Members of class 4 show the largest preference for reductions in power cuts and are against an increase in renewable energy. Further, they are in favor of privatization, be it with private companies or with co-operatives. We call class 4 members counter green reformers or CountGreenRef. Another way to look at the results is interpreting WTP values. As we used dummy variables for the cost parameter, the interpretation is not straightforward. Therefore, we reverse the calculation by dividing the cost parameter by the other parameters. Then, the interpretation, exemplified by the CostSensCons (class 1) scheduled power cuts, is as follows: An increase of costs by one unit – in this case 20% – would have to be compensated with (−3.381/0.524) ∗ 15 ≈ −97 min of reduced scheduled power cuts. The currency is minutes rather than INR and we lose the advantage of having one currency for all attributes. However, we can observe quantitatively the impact of a tariff hike on the respective attribute. For members of the GreenCons (Class 2) this required reduction is 21 min and for the CountGreenRef (Class 4) the required reduction amounts to 12 min. This means that the highest WTP for reduction of power cuts is borne by the CountGreenRef because they are willing to accept a tariff hike by 20% if the power cuts are reduced by only 12 min. For unscheduled power cuts the structure is similar but the WTP values are less in all classes. This analysis can be carried out further considering a tariff hike of 10%. Table 6 summarizes the


97

Table 6 Class interpretation and inverse WTP values. Class

CostSensCons (Class 1) GreenCons (Class 2) IndecLaym (Class 3) CountGreenRef (Class 4)

Name

Cost-sensitive conservatives Green conservatives Indecisive laymen Counter green reformers

Description

SCHED

Low WTP, prefer governmental supplier and green energy Green energy, governmental supplier, small tariff increases do not matter No consistent or meaningful preferences High WTP, against green energy, pro privatization

UNSCH

10%

20%

10%

20%

48 0 (–) 7

97 21 (–) 12

116 0 (–) 14

232 70 (–) 26

The last four columns give the inverse WTP values for a tariff increase of 10% and 20% for scheduled and unscheduled power cuts. The figure is the threshold reduction of power cuts in minutes that is required for an individual in the respective class to agree to a tariff hike.

classes and their inverse WTP values for reduction of scheduled and unscheduled power cuts.

Socio-demographic variables and class assignment In this section, we briefly investigate the relationship between class membership and socio-demographic variables. To do so, we applied a multinomial logit model with the (modal) preference classes as the dependent variables. This method has the disadvantage that it does not consider class probabilities but is straightforward in its interpretation and should suffice the goal to give a broad overview.15 Table 2 summarizes the socio-demographic variables which were used in the analysis. In the multinomial logit model, which we applied here, the IndecLaym (Class 3) serves as the base category because it distinguishes from the other classes by having no clear preferences. Hence, we can straightforwardly interpret the effects of the covariates relative to this class. We used LNINCPERHEAD and LNELECTRICITYEXP which are the logarithmic values of INCPERHEAD and ELECTRICITYEXP and AGE and AGESQ to capture non-linear effects. Table 7 summarizes the results and reports the multinomial odds ratios (relative risk ratios) i.e. eβ instead of β where β is the parameter estimate. Odds ratios indicate the multiplier of the odds being member of the IndecLaym relatively to the other classes. The overall model is highly significant with a chi square value with 27 degrees of freedom of 72.87. McFadden's Pseudo R2 is relatively low with 0.054. The odds ratio for HHMEMBERS for the CostSensCons (class 1) is 0.738. An odds ratio below one translates to a β value below zero. So, an increase of HHMEMBERS by one more person multiplies the odds of being member of the CostSensCons rather than the IndecLaym by 0.738. The same relationship is true for the GreenCons (class 2) and the CountGreenRef (class 4). A similar pattern can be observed for CATEGORY, which is dummy coded. Slum is chosen as the base, and when one changes from slum to middle class and high class, the odds increase accordingly. A respondent from the high class is more likely to be member in the other preference classes, especially the GreenCons (class 2) and the CountGreenRef (class 4). The odds ratio for the CountGreenRef for high class is 3.92, so the chances being member of CountGreenRef are nearly four times higher for high class respondents than for slum respondents. BACKUP shows odds ratios below one. This result can mean that respondents who have backup systems are more indifferent of power quality from the grid and hence have reduced interest in improvements. Women are more likely to be member of the IndecLaym than men yet the odds ratios are close to one. Age has a ushaped effect. Very young and very old people are less likely to be member of the IndecLaym relative to the other classes. The odds ratio for LNINCOMEPERHEAD is below one for all classes i.e. a higher per head income decreases the odds to be member of the CostSensCons, the GreenCons and the CountGreenRef relative to the IndecLaym. The opposite is true for LNELECTRICITYEXP. Here, a higher expenditure leads to a

15 A detailed discussion on different approaches to link socio-economic variables to latent classes can be found in Clark and Muthén (2009).

higher probability to be member of the CostSensCons and the CountGreenRef but not of the GreenCons. The analysis reveals significant effects on the probability of membership of the IndecLaym, in relation to the other classes only for some variables.16 Household size has a negative impact on the probability to be a member of the CostSensCons (class 1), the GreenCons (class 2) and the CountGreenRef (class 4) relative to IndecLaym (class 3), whereas respondents who belong to middle or high class are more likely to be part of the GreenCons. Moreover, increasing per head income is not enhancing the probability of being a member of the GreenCons and the CountGreenRef. Finally, sociodemographic variables explain only partially the heterogeneity of preferences for power service quality. Summary and conclusions In this paper, we analyzed data from a DCE to investigate preferences for improved electricity supply quality for private households in the emerging megacity Hyderabad in India. We applied a SALC logit model to incorporate discrete unobserved heterogeneity and to control for variance scale heterogeneity. We examined four criteria of power quality, namely, availability of electricity, external effects of power generation, organization of the power sector and the costs for improved electricity supply quality. The results indicated a large amount of unobserved preference heterogeneity and that about 90% of the population are not interested in the topic at all or not willing to pay for improvements in quality. In order to illustrate the results, we examined the estimated preference classes in more detail. We found two classes, the GreenCons and the CountGreenRef, whose members are willing to pay for improvements of power quality, yet their share in the population is less than 10%. The CostSensCons, the largest class in the sample, has preferences for an increase in the share of renewable energy and reduction of power cuts, yet is not willing to pay for it. The last class, the IndecLaym, constituted of respondents who did not understand the DCE or did not care and hence stated no clear preferences. Our results show further that most consumers are not interested in privatization of the power sector. The GreenCons and the CostSensCons prefer the current system with government-owned distribution companies. The analysis of the interaction of socio-demographic variables and class membership revealed a limited capability of such data to explain the heterogeneity. As expected, it turned out that respondents from higher social categories, living in smaller households, having higher electricity expenditures and no backup systems, tend to care more about their power quality. The analysis might not seem very promising for policy makers because the share of the reform orientated consumers is rather low. Our results show low WTP values for the majority of Hyderabad's domestic households. This means that improvements in reliability and in increases in renewable energies cannot be financed by increased domestic tariffs, at least not without reducing social welfare. A detailed cost–benefit analysis would go beyond the scope of this paper, yet the hypothesis 16 Changing the base alternatives and base categories lead to different p-values of the parameters. Hence, an interpretation of the significance levels has not much relevance.

98


Table 7 Estimation results: multinomial logit model. Class 1 AGE AGESQ HHMEMBERS HIGH CLASS MIDDLE CLASS SEX BACKUP LNINCPERHEAD LNELECTRICITYEXP N Number parameters Log-likelihood Pseudo R2 Chi squared

0.830 (−1.73) 1.002 (1.67) 0.738⁎⁎ (−2.74) 1.039 (0.04) 1.731 (1.20) 1.027 (0.08) 0.591 (−0.93) 0.421⁎ (−2.48) 1.537 (1.19) 687 30 −634.436 0.0543 72.87

2

4

0.977 (−0.21) 1.000 (0.23) 0.807 (−1.94) 4.249 (1.43) 3.315⁎⁎ (2.58) 1.425 (0.99) 0.727 (−0.55) 0.372⁎⁎ (−2.77) 0.997 (−0.01)

0.869 (−0.93) 1.002 (0.92) 0.859 (−0.93) 3.915 (0.90) 2.989 (1.55) 1.352 (0.54) 0.656 (−0.50) 0.483 (−1.36) 1.867 (1.10)

Exponentiated coefficients; t statistics in parentheses. ⁎p b 0.05, ⁎⁎p b 0.01, ⁎⁎⁎p b 0.001.

that the marginal benefits were larger than the marginal costs of such improvements does not seem realistic. However, our analysis also revealed that a minority has rather high WTP values for such improvements. Identifying them could provide at least some scope for financing improvements through domestic consumers. As we found that many respondents were not aware, not interested, or not capable of understanding the concept of renewable energies and the sector in general, governmental campaigns to promote these and other aspects could help in creating more awareness and hence more willingness to deal with the subject. This might result in a larger WTP for improvements considering the findings of our study. In general, we suggest policy makers in India to allocate more resources to consumer research and incorporate such findings into upcoming policy decisions. We argue that future policy decisions should incorporate stated preferences data. It is hardly possible to conduct cost benefit analyses with the available revealed preferences data because it does not reflect, first, the true valuations of the consumers and, second, does not have much indication for the valuation of different criteria of power quality. Stated preferences methods can be used to bridge this gap and help to absorb the existing WTP in tariff designs. Especially against the background of the ambitious aims of the Indian government to reduce CO2 emissions, contributions from the consumers become increasingly important. Further research is required to better understand consumer preferences including industrial and commercial users, and to compare marginal costs of increases in renewable energy and reduced power cuts with marginal benefits. Acknowledgments We would like to thank Philip Kumar and Vamsi Krishna and their team for their tremendous efforts during the field phase. Further thanks go to Markus Hanisch, Christian Kimmich, and Kaushik Deb and his team from TERI for supporting us during the pre-field phase. Special thanks goes to K Raghu from the Andhra Pradesh Transmission Cooperation Ltd., Rao Chelikani from the Tarnaka Residents Welfare Association, Hyderabad and Dr. Rama Rao from the Andhra Pradesh Electricity Regulatory Commission for valuable inputs on the

specification of the attributes used in the choice experiment. We highly appreciate advice on the Scale Adjusted Latent Class model from Terry Flynn from the Centre for the Study of Choice, Sydney. Further, we thank three anonymous reviewers for helpful comments. Financial support from the German Ministry for Education and Research (FKZ 01LG0506A1), Germany for the “Sustainable Hyderabad” project is gratefully acknowledged. References Abdullah S, Mariel P. Choice experiment study on the willingness to pay to improve electricity services. Energy Policy 2010;38(8):4570–81. [August, URL http://ideas.repec. org/a/eee/enepol/v38y2010i8p4570-4581.html]. Alpizar F, Carlsson F, Martinsson P. Using choice experiments for non-market valuation. Working Papers in Economics, 52. University of Gothenburg, Department of Economics; 2001. [URL http://EconPapers.repec.org/RePEc:hhs:gunwpe:0052]. Amador FJ, González RM, Ramos-Real FJ. Supplier choice and {WTP} for electricity attributes in an emerging market: the role of perceived past experience, environmental concern and energy saving behavior. Energy Econ 2013;40:953–66. [URL http:// www.sciencedirect.com/science/article/pii/S014098831300120]. Andhra Pradesh Electricity Regulatory Commission. Renewable power purchase obligation regulations. URL https://www.recregistryindia.in/pdf/RPO/AndraRECRegulation.pdf, 2012. Bujosa A, Riera A, Hicks R. Combining discrete and continuous representations of preference heterogeneity: a latent class approach. Environ Resour Econ 2010; 47:477–93. http://dx.doi.org/10.1007/s10640-010-9389-y. [URL http://dx.doi. org/10.1007/s10640-010-9389-y]. Burke PF, Burton C, Huybers T, Islam T, Louviere JJ, Wise C. The scale-adjusted latent class model: application to museum visitation. Tour Anal 2010;15:147–65. Campbell D, Doherty E, Hynes S, van Rensburg TM. Combining discrete and continuous mixing approaches to accommodate heterogeneity in price sensitivities in environmental choice analysis. 84th Annual Conference, March 29–31, 2010, Edinburgh, Scotland 91723. Agricultural Economics Society; 2010. [URL http://EconPapers. repec.org/RePEc:ags:aesc10:91723]. Carlsson F, Martinsson P. Does it matter when a power outage occurs? A choice experiment study on the willingness to pay to avoid power outages. Energy Econ 2008; 30(3):1232–45. Central Electricity Authority, Ministry of Power, G. o. I. Reliability indices 2009; 2009. Central Electricity Authority, Ministry of Power, G. o. I. Load generation balance report; 2011. Central Electricity Authority, Ministry of Power, G. o. I. Annual report 2012–13; 2013. Central Electricity Authority, Ministry of Power, G. o. I. CO2 baseline database for the Indian power sector. URL http://www.cea.nic.in/reports/planning/cdm_co2/database_9.zip, 2014. Central Electricity Authority, Ministry of Power, G. o. I. Co2 baseline database for the Indian power sector — user guide version 9.0. URL http://www.cea.nic.in/reports/ planning/cdm_co2/user_guide_ver9.pdf, 2014. Central Electricity Authority, Ministry of Power, G. o. I. Monthly all India installed generation capacity report February 2014. URL http://www.cea.nic.in/reports/monthly/ inst_capacity/feb14.pdf, 2014. CESS. Approach to the 12th five year plan of Andhra Pradesh. URL http://www.cess.ac.in/ cesshome/pdf/Draft_Approach_to_12th_Plan_for_discussion.pdf, 2013. Clark SL, Muthén B. Relating latent class analysis results to variables not included in the analysis; 2009. unpublished. Deb K, Garg A, Rommel K. Energy management for the emerging megacity Hyderabad. Bremen: Europäischer Hochschulverlag EHV; 2012. Fiebig DG, Keane MP, Louviere J, Wasi N. The generalized multinomial logit model: accounting for scale and coefficient heterogeneity. Mark Sci 2010;29:393–421. [May]. Flynn TN, Louviere JJ, Peters TJ, Coast J. Using discrete choice experiments to understand preferences for quality of life. variance-scale heterogeneity matters. Soc Sci Med 2010;70(12):1957–65. [URL http://www.sciencedirect.com/science/article/pii/ S0277953610002352]. Fosgerau M, Hess S. A comparison of methods for representing random taste heterogeneity in discrete choice models. Eur Transp-Trasp Eur 2009;42:1–25. Greene WH, Hensher DA. A latent class model for discrete choice analysis: contrasts with mixed logit. Transp Res B Methodol 2003;37(8):681–98. Hanisch M, Kimmich C, Rommel J, Sagebiel J. Coping with power scarcity in an emerging megacity: a consumers' perspective from Hyderabad. Int J Glob Energy Issues 2010; 33(3&4):189–204. Hensher D, Greene W. The mixed logit model: the state of practice. Transportation 2003; 30(2):133–76. [URL http://dx.doi.org/10.1023/A:1022558715350]. Hensher DA, Rose JM, Greene WH. Applied choice analysis: a primer. Cambridge University Press; 2005. Hess S, Beharry-Borg N. Accounting for latent attitudes in willingness-to-pay studies: the case of coastal water quality improvements in Tobago. Environ Resour Econ 2012:1–23. Hess S, Rose J. Can scale and coefficient heterogeneity be separated in random coefficients models? Transportation 2012:1–15. Lal S. Can good economics ever be good politics? Case study of the power sector in India. World Bank working paper, vol. 83. Washington and DC: World Bank; 2006. [URL http://siteresources.worldbank.org/EXTDEVCOMMENG/Resources/ finalpowersectorindiaebook.pdf].

J. Sagebiel, K. Rommel / Energy for Sustainable Development 21 (2014) 89–99 Lancaster KJ. A new approach to consumer theory. J Polit Econ 1966;74:132. [URL http:// ideas.repec.org/a/ucp/jpolec/v74y1966p132.html]. Magidson JK, Vermunt JK. Removing the scale factor confound in multinomial logit choice models to obtain better estimates of preference. Sawtooth Software Conference Proceedings. Sawtooth Software; 2008. p. 139–55. Manski CF. The structure of random utility models. Theory Decis 1977;8(3):229–54. McCutcheon AL. Latent class analysis. In: McCutcheon Allan L, editor. 4th ed. Newbury Park and London and New Delhi: Sage; 1987. McFadden D. Conditional logit analysis of qualitative choice behavior. In: Zarembka P, editor. Frontiers in econometrics. Academic Press; 1974. p. 105–42. Mueller J, Rommel J. Is there a future role for urban electricity cooperatives? The case of Greenpeace energy. In: Ramos-Martín J, Giampietro M, Ulgiati S, Bukkens S, editors. Can we break the addiction to fossil energy? Barcelona: Universitat Autònoma de Barcelona; 2011. p. 185–95. Reddy MT, Raghu K. Favouring the rich. Econ Pol Wkly 2012;47(33). [Aug]. Revelt D, Train K. Mixed logit with repeated choices: households' choices of appliance efficiency level. Rev Econ Stat 1998;80(4):647–57. [URL http://www.mitpressjournals. org/doi/abs/10.1162/003465398557735].

99

Rosen S. Hedonic prices and implicit markets: product differentiation in pure competition. J Polit Econ 1974;82(1):34–55. Sagebiel J, Müller JR, Rommel J. Are consumers willing to pay more for electricity from cooperatives? Results from an online Choice Experiment in Germany. Energy Res Social Sci 2014;2:90–101. http://dx.doi.org/10.1016/j.erss.2014.04.003. Sreekumar N, Thimma Reddy M, Raghu K. Strengths and challenges of Andhra Pradesh power sector. Econ Pol Wkly 2007;42(45/46):24–7. Thurstone LL. A law of comparative judgment. Psychol Rev 1927;34:273–86. Tongia R. The political economy of Indian power sector reforms. In: Victor DG, Heller TC, editors. The political economy of power sector reform. Cambridge: Cambridge Univ. Press; 2007. p. 109–74. [URL http://pesd.stanford.edu/publications/political_economy_of_indian_power_sector_reforms_the/]. Vermunt JK, Magidson JK. Latent gold 4.5: choice, syntax module. URL http:// statisticalinnovations.com/products/lgchoice.html, 2005.