Do Flemish Households Value Renewables? - TU Dresden

3 downloads 118 Views 201KB Size Report
Mar 31, 2011 - Choice experiment, Renewables, Conditional Logit model,. Random Parameter logit model, Willingness-to-pay. JEL-classification: C25, C93 ...
Do Flemish Households Value Renewables?

Guido Pepermans

31 March 2011 (Draft version – do not quote without permission of the author)

Abstract This paper assesses whether and to what extent Flemish households are willing to pay for renewable power supply. Via a choice experiment, households were offered the choice between as set of green electricity contracts, characterised by the renewables share in electricity supplied to their dwelling, the renewable generation technology (wind, biomass, PV…) and its impact on the electricity bill. A main effects conditional logit model and a main effects random parameter logit model are estimated. The estimation results show that households prefer ‘green contacts’, but renewables technologies are valued differently. The estimates are then used to assess the marginal willingness to pay by Flemish households for each of the contract attributes. From a policy perspective, the results suggest that a not too small proportion of Flemish households would be willing to switch to another power supply contract if that can be done at limited cost. Moreover, the results suggest that it would not be a good idea to focus on the deployment of one or only a few technologies. Policies resulting in a diversified portfolio of technologies are positively valued by households and will obtain broader support. Keywords:

Choice experiment, Renewables, Conditional Logit model, Random Parameter logit model, Willingness-to-pay.

JEL-classification:

C25, C93, D12, Q41

Corresponding Address:

Guido Pepermans Faculty of Economics and Management HUBrussel Stormstraat 2 1000 Brussels Belgium e-mail: [email protected]

2

Do Flemish Households Value Renewables? In 2007, European leaders have committed Europe to increase energy efficiency and move towards a low-carbon economy. In order to achieve this, three climate and energy targets were set, to be met by 2020. First, greenhouse gas emissions should be reduced with 20% below the 1990 levels. Second, 20% of the EU’s energy consumption should be covered with renewable resources, and, third, primary energy use must be reduced with 20% below the projected levels by improving energy efficiency. These targets are collectively known as the 20-20-20 targets and in June 2009 they were transformed into binding EU legislation. These targets were also translated into binding targets for the Member States. For Belgium, the renewables target has been set at 13%, to be achieved in 20201. The country-level GHG emission reduction and energy efficiency improvement targets have not been set yet, but it is likely that, whatever these targets will be, they will require using properly designed policy measures2. Designing these policy measures is not an easy task and more stringent targets – likely to be set in the future – will make it even more difficult. Regarding GHG emissions, the EU has already decided to submit the energy intensive sectors to an EU-wide Emission Trading Scheme (EUETS), but apart from that, no further policy measures have currently been decided upon. It is largely left to the Member States to decide what policy measures to take. Regarding renewables, two approaches can be thought of to stimulate its use: the government can collect and spend funds for renewable energy, or the private sector can be asked or mandated to do this (Wiser (2007)). Concerning the payment vehicle used to finance the green electricity policy, one can distinguish a voluntary and a mandatory approach. With the voluntary

1

2

Directive 2009/28/EC of the European Parliament and of the Council of 23 April 2009 on the promotion of the use of energy from renewable sources and amending and subsequently repealing Directives 2001/77/EC and 2003/30/EC. One exception is that, regarding the sectors not covered by the EU ETS (such as transport, housing, agriculture and waste), an effort sharing decision has been taken at the EU level. For Belgium, it has been decided that, by 2020, GHG emissions should be reduced with 15% below the 2005 emission levels.

3 approach, it would be left to customers to decide how much green electricity to buy and thus to decide how much to spend. With the mandatory approach, the cost of the stimulus program would be socialized via a surcharge on the electricity bill. Along the lines of Wiser (2007), one can then in principle distinguish 4 types of policy: government provision of green electricity with mandatory payments, government provision with voluntary contributions, private provision with mandatory contributions and private provision with voluntary contributions. Except for the second type – government provision based on voluntary contributions – all combinations have been used in practice to stimulate green electricity. For example, a tradable green certificates (TGC) mechanism is a private provision with mandatory payments policy. The purpose of this paper is to investigate whether it would be feasible – on top of the existing TGC support mechanism – to further stimulate renewables via a private sector initiative based on a voluntary payment mechanism. To be more precise, this paper investigates whether there is a market potential for green electricity contracts in Flanders. The paper will provide useful information to the private sector, i.e. to retail companies planning to market green electricity as any other private good or service. In addition, the results will also provide useful insights to policy makers. Based on the estimated models, they will have a view on the relative preferences of households towards the various existing technologies to generate green electricity. In summary, this paper looks at the responsiveness of households for marketing polices aimed at stimulating the use of green electricity3. The purpose is to find out whether it would be worthwhile from a public policy perspective to fine-tune existing policies in order to increase their effectiveness and the cost-efficiency. A stated choice approach will be used to do this, i.e. we use a small choice experiment carried out in Flanders in the first semester of 2010 as part of two master thesis projects. For reasons explained below, we feel that, despite the limited

3

In this paper, green electricity is used as short notation for electricity produced by generation technologies using (a mix of) renewables. Grey electricity is used to indicate the current generation mix, thus also including a small share of electricity generation based on renewables.

4 number of respondents, the conclusions are valuable and useful for both public and private policy making. With ‘fine-tuning’, we are thinking of additional efforts that the private sector can make on top of the already existing quota mechanism, to market green electricity. In principle, marketing green electricity can be realized relatively easy by using certificates of origin as an instrument to prove the green origin of the generated green electricity. However, in order to do this efficiently it is necessary to have sufficient information about household preferences for electricity generated by using renewables. The choice experiment focuses on the question whether households are – on a voluntarily basis – willing-to-pay (WTP) a premium for green electricity on top of their current per kWh price (already including a surcharge to cover the cost of the Flemish TGC mechanism) for grey electricity. Furthermore, we also focus on two additional questions: i) do households have a different willing-to-pay for different technologies and, ii) can household preferences regarding RES technologies be considered to be homogeneous? In section 1 we will briefly survey some relevant literature on WTP estimation. Section 2 introduces the theoretical background for the methodology and section 3 then describes the data and the applications in the Flemish household sector. Finally, section 4 concludes.

1.

LITERATURE REVIEW

Finding the willingness-to-pay for renewables requires information about people’s preferences. Whereas in the past, revealed preference approaches were used more frequently, we now observe that, for many applications, the focus has shifted to using stated preference techniques to collect data, i.e. data are collected via surveys in which respondents are asked to assess hypothetical situations. This is also the approach followed in this paper. Essentially, two stated preference techniques are available to assess WTP values: the contingent valuation technique and the choice experiment. This section provides a non-exhaustive survey of the recent literature on the use of stated preference techniques in the field of energy economics.

5 The WTP for green electricity Investigating the WTP for green electricity via stated choice techniques has been the topic of a number of papers. Most of the older applications used a contingent valuation approach, while, more recently, a number of choice experiments (CE) have been conducted. In this nonexhaustive survey, we focus on papers using CE approaches, but example of contingent valuation studies in the field of renewables are Nomura and Akai (2004), Wiser (2007), Yoo and Kwak (2009) and Zarnikau (2003). Irrespective of the approach, most studies report a positive WTP for various characteristics of green electricity. Some studies define green energy as a generic product described in very general terms. For example, Zarnikau (2003) uses a contingent valuation approach to study the end-user’s WTP for utility investments in renewable energy and energy efficiency. Other authors, such as Bergmann, et al. (2006) and Longo, et al. (2008), also take a more global point of view but use choice experiments to collect data. Bergmann, et al. (2006) defines green energy in terms of its environmental attributes, such as, for example, landscape impact, wildlife impact and air pollution. They find that households are quite sensitive to projects that have a large impact on landscape, but much less so when the landscape impact is considered moderate to low. Wildlife effects are also valued highly as well as avoiding air pollution. Overall, in their study preferences do not seem to vary with income levels, but geographical characteristics do seem to play a role when valuing some of the attributes. Longo, et al. (2008) set up a choice experiment in which four potential effects of a renewables policy are being considered: GHG emission reductions, short term security of supply (blackouts), employment effects and the price impact. Their results show that respondents are not indifferent to policies reducing GHG emissions, creating jobs and reducing blackouts. Scarpa and Willis (2010) investigate the households’ WTP for renewable micro electricity generation technologies in the UK. They conclude that, while renewable energy adoption is

6 significantly valued by households, this value is – for a large majority of households – insufficient to cover the higher capital cost of these technologies. Finally, other authors, such as Borchers, et al. (2007), use choice experiments to focus on the input side of green electricity rather than on the output side. They estimate the WTP for different renewable technologies and conclude that households have a positive WTP for renewables with a relative preference for solar energy over wind, biomass, farm methane and a generic green mix. This paper will use a similar approach. Similar exercises in the field of energy economics The choice experiment technique can be used to value many different aspects of energy policy, such as for example the WTP for energy efficiency measures or the WTP for continuous power supply. Applications in the area of energy efficiency are less numerous than those in the area of renewable energy, but, in general, they all arrive at the same conclusion, i.e. households attach positive value to energy efficiency measures. Revelt and Train (1998) estimate a random parameter (RPL) model. Such a model can be considered an extension of the conditional logit (CL) model in which heterogeneous consumer preferences are allowed for. These authors collected data via a choice experiment on households’ choices of appliance efficiency levels. To be more precise, a sample of respondents was presented a number of choice sets in which it was offered the choice between 2 or 3 refrigerators with different efficiency levels, each supported either via a rebate, a loan or no support at all. Customers were then asked which appliance they would choose. Banfi, et al. (2008) focus on the WTP of households, either owners or tenants, for air renewal systems and improved window and facade insulation. Two samples are taken, one for house owners and one for tenants. Based on an estimated logit model with fixed effects, these authors conclude that the benefits from energy efficiency applications are significantly valued by households, irrespective of the ownership status.

7 Shen and Saijo (2009) use a choice experiment approach to assess the impact of energy efficiency labels on the consumer’s WTP in Shanghai. The study focuses on consumers’ choices of air conditioners and refrigerators. Their results suggest that Chinese customers are well aware of the existence of the energy efficiency label and that it is positively valued. Moreover, it was found that the WTP for energy efficient refrigerators was higher than that for more energy efficient air conditioners, suggesting that the Shanghai population has greater incentives to pay more for appliances that are being used more frequently. Another field of energy economics in which stated preference techniques and more specifically choice experiments are often used, is short term security of supply or power outages. Applications can be found in Beenstock, et al. (1998), Carlsson and Martinsson (2008) and Pepermans (2010). In all cases, power outages are described in terms of their main characteristics, such as frequency and duration of the outages. Respondents are then asked to indicate their most preferred power outage profile within a number of choice sets shown to them. The results suggest that a large fraction of households is WTP for uninterrupted power supply, but at the same time, a non-negligible fraction is also willing to accept more power outages if that would be compensated via a lower electricity bill. The following section briefly introduces the choice experiment methodology, some relevant literature and the techniques used in this paper to estimate the preference structure.

2.

METHODOLOGY

2.1.

The Choice Experiment

The basic idea of a choice experiment is quite simple: face households with a set of hypothetical items (goods, services, options, situations…), each of which is described by a number of typical characteristics or attributes. Within this set, they then have to indicate the item they prefer. These stated choices reveal information about their preferences, which can then be used to assess the relative value of the different attributes describing the items. Thus, at the basis of the choice experiment approach lies the idea that any good or service can be described by a number

8 4

of characteristics or attributes and by the levels these attributes take . The approach comes close to a real market situation as respondents effectively have to make choices and one of the major advantages is that, by describing a product or a service in terms of attributes and attribute levels, the relative importance (value) of these attributes can be assessed. In this paper, the service we consider is a ‘green electricity’ retail contract, i.e. a contract to supply electricity based on renewables. On the basis of a brief literature survey and observed practice in Flanders, we identified a number of attributes describing green electricity: the technology used to generate the electricity, the share of electricity supplied originating from the renewable source, and the impact of this green contract on the annual electricity bill. These attributes and the values they can take in the choice experiment are summarized in Table 1. Attribute Technology

Percentage Green in contract

Percentage impact on Invoice

Levels Onshore wind Offshore wind Biomass Large scale PV Mix of wind, large scale PV and Biomass Mix of wind and large scale PV 33% 67% 100% -5% 0% +10%

Nr of levels 6

3

3

Table 1: Attributes and levels in the green electricity choice experiment. The choice experiment based on the attributes and levels in Table 1 has a full factorial design of

6 × 3 × 3 = 54 profiles or contracts. Clearly, the number of contracts would increase dramatically when the number of attributes and/or levels would increase. Obviously, in that case it would become too complicated for respondents to evaluate all available profiles. However, a subset of profiles – a fractional factorial design – can be constructed such that – with some loss of information – the most important and relevant effects can be estimated. Whereas a full factorial design would allow unbiased estimation of all possible so-called main and interaction effects in a linear model, a fractional factorial design will only allow estimating some of these effects, probably in a biased way. Which effects can be estimated and which ones will be biased will

4

We refer to Bateman, et al. (2002) and Amaya-Amaya, et al. (2008) for a more elaborate discussion of the CE technique.

9 depend on the constructed fractional factorial design (Louvière, et al. (2003)). The exercise presented in this paper is based on a fractional factorial design that allows estimating all main effects while giving maximal consideration to balancing and orthogonality5. The selected profiles or contracts are then combined in choice sets. In addition to these profiles, a so-called status quo or opt-out option is also included. This allows respondents to decide not to purchase or indicate any of the presented green electricity contracts, but rather to stick to their current situation. In that way, respondents can always compare the offered profiles with their current situation. Depending on the difficulty of the evaluation process, more or less choice sets should be presented to respondents. In our experiment, respondents were asked to evaluate 18 choice sets in total. This is a large number, but a small pre-test showed that it was feasible and not too burdensome for respondents. The possibility to present a number of choice sets is considered an advantage of using choice experiments as multiple observations per respondent can be obtained in that way.

2.2.

Theoretical background

To assess the WTP for green electricity we estimate a discrete choice model. Conjoint choice data are typically analyzed with logit models, see for example Beenstock, et al. (1998). As Train (2003) and Moore (2008) point out, one drawback of the standard logit model is that homogenous preferences are assumed. Logit models tackle heterogeneity only to the extent that it is explained by interacting demographic or household characteristics with product attributes. An alternative approach to allow for heterogeneous preferences is to estimate a Random Parameters or Mixed Logit Model (Train (2003)). Over the past years, the number of applications of RPL models has increased, e.g. Carlsson and Martinsson (2008) and Pepermans (2010) for preferences for continuous power supply in Sweden and Flanders, respectively, and

5

According to Louviere (1988), main effects explain between 70% and 90% of respondent behaviour. The efficiency of the fractional factorial designs has not been evaluated. However, for the construction of the designs, we used the methodology described in Street, et al. (2008).

10 Revelt and Train (1998) for an application in the area of energy efficiency. As explained by Moore (2008) one drawback of the standard RPL model is that, although heterogeneous preferences are allowed for, the sources of heterogeneity often remain unexplored. This could be handled by estimating the distribution of the random preference parameters conditional upon individual characteristics. See E. R. Morey, et al. (2003) and Moore (2008) for illustrations of this approach. Alternatively, heterogeneous preferences could also be introduced via a latent class model along the lines of E. Morey, et al. (2006) or Boxall and Adamowicz (2002). E. Morey, et al. (2006) considers group or class membership as exogenous, whereas Boxall and Adamowicz (2002) assume group membership to be endogenous. Moore (2008) compares the three modeling approaches and concludes that assuming heterogeneous preferences adds to the explanatory power of the models. Furthermore, he finds that, despite differences in the underlying assumptions and in the parameter estimates, the WTP-estimates derived from the three models show little difference. Therefore, from a policy perspective the main message is that it does not matter how preference heterogeneity is included in empirical models, as long as it is included. Random Utility Model We use the Random Utility Model (RUM) to analyse household preferences. The RUM is based on random utility theory which starts from the assumption that decision units maximize utility, i.e. when a decision maker is faced with a set of different alternatives, he or she will always choose the one with the highest utility. Let decision unit n face T consecutive choice problems each of which implies a choice to be made between J alternatives. From each of the alternatives

j a utility level Unjt can be obtained, which is known to the decision unit but is only partially observed by the researcher, i.e.

Unjt = Vnjt + ε njt ,

(1)

11 with Vnjt observed utility and ε njt unobserved utility, represented as a random term. For each choice problem C , a decision unit will select the alternative that provides maximal utility. Thus, at time t alternative j is chosen by decision unit n when

Unjt > Unit

∀i ≠ j ∈C

(2)

Due to the presence of the random component in equation (1) only probabilistic statements can be made about the respondent’s choices. The probability of choosing alternative j from choice set C can be written as P ( j C ) = P ( Unjt > Unit , ∀i ≠ j ∈ C ) = P (Vnjt + ε njt > Vnit + ε nit , ∀i ≠ j ∈ C ) = P ( ε nit < ε njt + Vnjt − Vnit , ∀i ≠ j ∈ C )

(3)

Assume that ε njt is i.i.d. type I extreme value. It can then be shown that, for decision unit n , the probability of choosing j , when faced with choice set C at time t , equals V

Pnjt =

e njt . ∑ eVnit

(4)

i∈C

Usually, observed utility is assumed to be linear in the parameters, i.e.

Vnjt = β xnjt ,

(5)

with xnjt a vector of alternative-specific attributes and β the vector of parameters to be estimated. Equations (4) and (5) define the conditional logit model. Note that the parameter vector is not indexed, implying that preferences are assumed to be homogeneous. This is a rather extreme assumption that can be relaxed by allowing tastes to vary over the population with density f ( β ) (Train (2003)). We can then rewrite equation (5) as

Vnjt = βn xnjt ,

(6)

12 where the heterogeneity of preferences is now made explicit by indexing β . Note that β n is assumed constant over time, i.e. preferences of decision unit n are stable over consecutive choice situations, which for the current application is a realistic assumption. In this paper, we assume preferences to follow a normal density, i.e. β n = µ β + ηn with

β n ∼ N ( µ β , σ β ) or ηn ∼ N ( 0,σ η ) , where η n is a vector of individual-specific deviations that are assumed to be normally distributed with mean 0 and standard deviation σ η . A decision unit knows his own β n when choosing an alternative, but the researcher does not. Conditional upon

β n , the probability of decision unit n choosing alternative j at time t is Lnjt ( β n ) =

e

β n xnjt

∑ eβ

n xnit

.

(7)

i∈C

Knowing that ε njt is i.i.d. extreme value over decision units, alternatives and time, we can write the conditional probability that a decision maker will make a given sequence of choice j = { j1 , j2 ,..., jT } as T

Lnj ( βn ) = ∏ t =1

e

β n x njt t

∑e

β n xnitt

i∈C

As the researcher does not know β n , he or she has to consider all possible values of β to arrive at the unconditional choice probability of decision unit n choosing the sequence of alternatives j : Pn j = ∫ Lnj ( β ) f ( β | µβ , Σ β ) d β

or T

Pn j = ∫ ∏ t =1

e

β x njtt

∑e i∈C

β x nitt

f ( β |µβ , Σ β ) dβ .

(8)

13 Equation (8) cannot be solved analytically, but simulation techniques can be used to solve for the preference parameters that maximize the simulated log-likelihood function6 (Train (2003)). Model specification Observed utility is specified as a linear function of the attributes of the contract for green k

electricity (share of renewables ( G ), renewables technology ( X ) and effect on the electricity bill ( B )). In its most general form, utility Unit for individual n of alternative i ( i = 1,2,3 ) in choice set t ( t = 1,...,18 ), is written as (the subscript n is omitted to simplify notation): Uit =

∑β

j =1,2

C j

ASC j + ( β G + η G ) Git +

∑ (β

k∈Tech

k

+ η k ) X itk + β B Bit + ε it

(9)

where ASC j is a dummy variable equal to 1 if i = j and zero otherwise7, G is a continuous variable with domain {0,1} , X k a dummy variable being equal to 1 if technology k is used and zero otherwise, and B a continuous variable expressing the bill impact in € per year of the described contract. Finally, the variables η i reflect the individual specific preference deviations ( i = G, k ). Although the choice experiment approach does not provide direct estimates of the WTP, it can be estimated indirectly via the estimated parameters of equation (9). In the next section we will discuss how this can be done.

2.3.

Estimating the Willingness-to-Pay

Once the model has been estimated, the results can be used to estimate household specific willingness to pay values. Under the assumption of a standard conditional logit model with observed utility linear in income (see eq. (5)) , the consumer surplus associated with a set of

6 7

Actual estimations were done with STATA’s mixlogit procedure.

ASC3 , the alternative specific constant for alternative 3 (the status quo), is left out in order to avoid multicollinearity problems.

14 alternatives takes a closed form that is easy to calculate (see also Train (2003)). The consumer surplus derived from the chosen alternative i is simply the utility derived from that alternative, expressed in monetary terms. Knowing that a decision maker chooses the alternative that maximizes his or her utility, the consumer surplus is 1

CSn =

βB

max (Un j )

(10)

j∈C

with β B representing the preference parameter related to the monetary attribute. Dividing by

β B translates utility in monetary terms. However, the researcher does not observe the utility Unj linked to the utility maximizing alternative. He only observes Vnj and he knows the distribution of the error term. Therefore, only expected consumer surplus can be calculated, i.e. E ( CSn ) =

(

E max (Vnj + ε nj )

1

βB

j∈C

)

(11)

McFadden (1973) and McFadden (1995) show that, if ε nj is i.i.d. extreme value and utility is linear in income (i.e. β B , the marginal utility of income, is constant), then this expression reduces to8

E ( CSn ) =

 V  ln ∑ e nj  + K , β  j∈C  1

B

(12)

with K a number known as Euler’s constant. An alternative interpretation of equation (12) is that E ( CSn ) is the average consumer surplus in the subpopulation of people who have the same representative utilities as consumer n . The total consumer surplus can then be calculated as the weighted sum of E ( CSn ) over a sample of decision makers, with the weights reflecting the

8

A more complex formulation of the change in consumer surplus is needed when the marginal utility of income is not constant. However, when marginal utility of income is constant over a range of income levels that correspond to the policy, then equation (12) can also be used (Train (2003), p. 61).

15 numbers of people in the population who face the same representative utilities as the sampled person (Yu (2003), p. 60). The change in consumer surplus that results from a change in the alternatives and/or the choice set is then equal to

∆E ( CSn ) =

1   V After ln ∑ e nj B   β   j∈C After

  V Before  − ln  ∑ e nj   j∈CBefore

     

(13)

When the purpose is to compare two alternatives or profiles, for example the base case (the status quo) and an altered case, and if both deterministic utility terms between accolades are linear in the attributes, then equation (13) reduces to ∆E ( CSn ) =

1

βB

{V

After n

− VnBefore }

(14)

If the purpose is to evaluate the change in one attribute and if deterministic utility is linear in the attributes, then equation (14) further reduces to the ratio of the marginal utility of the attribute and the marginal utility of income, also known as the marginal willingness to pay9. For models in which only main effects are estimated (as will be the case in this paper), the marginal willingness to pay for a change in a single quantitative attribute q is defined as

WTPq = −

3.

βq βB

(q = G, k )

(15)

ESTIMATION RESULTS

The analysis in this paper focuses on the Flemish residential sector and is based on a small survey containing 182 households. The Flemish region is the Dutch speaking northern part of Belgium, counting just over 6 million inhabitants and about 2,5 million households. The data

9

Champ, Boyle et al. (2003), p. 195-196.

16 10

were collected in the first semester of 2010 via a web survey . This approach was followed because of timing and cost considerations: web based surveys are faster to set up and fill-in, and less costly to organise11. The survey had three parts. The first part collects information on the respondent’s knowledge and attitudes regarding renewables. These questions help to prepare the respondent for the choice experiment, presented in the second part of the questionnaire. As discussed below, our choice experiment is kept fairly simple and straightforward and we therefore do not expect too much difficulties for respondents to answer the questions. Finally, in the third part of the survey we collect information on relevant socio-demographic characteristics (such as household size, education, income level…). Variable Name Sex Age distribution

Education

Household income

Annual electricity bill Province

Male Female ≤30: 31 to 50: >50: High school or less: Bachelor: Master: €3500: Unknown: Average: std. Deviation. Antwerpen: Limburg: Oost-Vlaanderen: Vlaams-Brabant: West-Vlaanderen:

71,5% 28,5% 32,6% 40,3% 27,1% 42,0% 31,7% 26,1% 27,8% 27,8% 27,1% 17,3% €713 €343,7 26,7% 20.6% 15,7% 15,8% 21,2%

Table 2:Some descriptive statistics. Cleaning up the responses resulted in a dataset containing 146 households that could be used to estimate the CL and the RPL models and the corresponding WTP values12. Table 2 summarizes some descriptive statistics for the cleaned database.

10

11

12

I am very grateful to Roel Claessens and Viktor Van Den Corput for helping with the preparation of the questionnaire and the collection and cleaning of the data. The data were collected as part of their master thesis. For a comparison and discussion of face-to-face and web surveys, we refer to Nielsen (2010). This author concludes that, for his study, the mean and median WTP values are statistically indistinguishable for both survey modes, despite the fact that response rates and other validation criteria can differ significantly. The initial sample contained 182 households. For some households, the data could however not be used for our analysis due to missing or inaccurate data, mainly regarding the size of the electricity bill.

17

3.1.

The conditional logit model

Several specifications have been tested for the CL model, but the relatively small sample makes it difficult to precisely estimate interaction effects with household specific characteristics. We therefore decided to integrate heterogeneous preferences by assuming that preferences follow a predefined distribution, the so-called mixed logit or random parameters model. The results obtained for this latter model will be discussed in section 3.2. Table 3 summarizes the estimation results for both models: a main effects conditional logit model and a main effects mixed logit model. As explained, neither of the models includes interaction effects or household covariates. Both models are estimated on the same set of 7884 observations, collected from the 146 households. As noted before, one advantage of the stated preference approach is that it allows collecting multiple observations per respondent. Respondents were asked to evaluate 18 choice sets, each containing 2 alternatives and the status quo and thus resulting in 18 x 3 x 146 = 7884 observations. Model specification The following specification was estimated, the results of which are summarized in Table 3: Uit =

∑β

j =1,2

C j

ASC j + β G Git +



β k X itk + β B Bit + ε it

(16)

k∈Tech

Two alternative-specific constants are added for the non-status quo alternatives. These constants capture variations in choices that cannot be explained by the attributes or by socio-demographic covariates included in the model. As alternatives are unlabelled, we expect the parameters of these constants to be equal within each model specification. However, as discussed by Champ, et al. (2003), these alternative-specific constants might capture a status quo bias and they might therefore be significantly different from zero. Positive values would indicate that the respondents prefers to move away from the status quo, a negative value would indicate conservative preferences from the part of the respondent. See Hartman, et al. (1991) and

18 Beenstock, et al. (1998) for an assessment of the status quo effect and its implications in the context of reliable electricity supply.

Variable Alt. Spec. Const 1 Alt. Spec. Const 2 Effect on Invoice Renewables % St. Dev. Onshore Wind St. Dev. Offshore Wind St. Dev. Biomass St. Dev. Large scale PV St. Dev. Mix Wind and PV St. Dev. Number of households Number of obs.: Wald χ2 (10 df): LR - χ2 Prob > χ2: Log pseudolikelihood:

Unit Dummy Dummy € %point

Condit. Logit Coef. Std. Err. 1,486*** 0,292 1,531*** 0,300 -0,008*** 0,001 1,084*** 0,177

Dummy

-0,323***

0,087

Dummy

-0,331**

0,112

Dummy

-0,793***

0,100

Dummy

-0,713***

0,119

Dummy

-0,128

0,099

146 7884 221,77 0,000 -2143,77

Rand. Param. Logit Coef. Std. Err. 3,906*** 0,261 4,194*** 0,265 -0,015*** 0,001 0,020*** 0,006 0,075*** 0,006 -0,569*** 0,163 1,195*** 0,181 -0,285 0,311 0,880*** 0,222 -1,578*** 0,201 1,717*** 0,225 -1,136** 0,348 1,781*** 0,210 0,043 0,304 0,446 0,238 146 7884

% posit. Pref.

60,5% 31,7% 37,3% 17,9% 26,2% 53,8%

1293,55 0,000 1496,99

legend: * p