The Environmental Impacts of Renewable Energy: An Investigation of Public Preferences
Nick Hanley1,* Ariel Bergmann and Yong-Joo Kim2
Abstract This paper estimates the magnitude of the external costs and benefits of renewable energy technologies in Scotland, a country that has set particularly ambitious targets for expanding renewable energy. The external effects considered are those on landscape quality, wildlife and air quality, whilst employment impacts are included as a fourth attribute. The methodology used is the Choice Experiment technique. We compare results from a simple Multi-Nomial Logit model with those from panel and non-panel versions of a Random Parameters Model. Welfare changes for different combinations of impacts associated with different renewable energy investments are calculated.
Keywords: renewable energy; external costs and benefits; choice experiments JEL classification: Q42, Q51
1. Economics Department, University of Stirling, Stirling FK9 4LA, Scotland, UK. Phone +44 1786 466410 Fax +44 1786 467469. Email
[email protected] 2. School of Architecture and Planning, University of Newcastle-upon-Tyne * corresponding author
1. Introduction Increasing the proportion of power derived from renewable energy sources is becoming an increasingly important part of the strategies of many countries to achieve reductions in greenhouse gas emissions (Owen, 2004). However, renewable energy investments can often have environmental impacts that need to be taken into account if socially-optimal investments are to be made. This paper attempts to estimate the magnitude of these impacts in Scotland, a country which has set particularly ambitious targets for expanding renewable energy. The environmental effects considered are those on landscape quality, wildlife and air quality. Unlike other papers in the literature, we do not restrict our investigation to the effects of particular technologies (Hanley and Nevin, 1999; Alvarez-Farizo and Hanley, 2001), but consider impacts applicable to a wide range of renewable technologies. The methodology used to do this is the Choice Experiment technique. Since preferences over environmental impacts likely vary significantly across the population, Random Parameter Logit models are estimated as well as Multi-Nomial Logit models. The renewable technologies considered include hydro, on-shore and off-shore wind power, and biomass. Welfare changes for different combinations of impacts associated with different renewable investment strategies are calculated. In what follows, Section 2 sets out some background detail on energy policy in Scotland. Section 3 provides an overview of the Choice Experiment method, focussing on the case where preferences are allowed to vary across respondents, whilst in Section 4 the design and conduct of the empirical study is outlined. Section 5 presents results from a simple MultiNomial Logit (MNL) model, in terms of parameter estimates and implicit prices: we also test there for differences between urban and rural households, and between low and high income households. We then investigate what differences occur in moving away from the MNL model to Random Parameter models, dependent on whether or not one allows for correlation between respondents' choices. Section 6 evaluates the welfare effects of alternative investment scenarios in renewables, whilst the final section presents some conclusions.
2
2. Renewable Energy Policy in Scotland The Scottish Executive has set two challenging targets for the production of electricity from renewable sources in the next 20 years. The first is that by 2010, 18 per cent of electricity consumed should come from renewable generation. The second is that by 2020, this portion should rise to 40 per cent (Scottish Executive, 2002a). Currently only 10% of electric energy produced in Scotland comes from renewable sources such as wind energy, hydro and waste-to-energy plants (Scottish Executive, 2002b). These renewable energy targets are ambitious by international standards: only Denmark has a higher target within European Union (de Vries et al, 2003). The major political reasons for promoting renewable energy are external to Scotland. The United Kingdom has accepted a legally binding target of reducing emissions of a bundle of greenhouse gases by 12.5% below 1990 emission levels by 2008-2012, as its share of the European Union negotiated target under the Kyoto Protocol. The UK Energy White Paper OurEnergy Future –Ccreating a Low Carbon Economy, published in February 2003, sets a much more ambitious domestic target of reducing CO2 emissions by some 60% of current levels by 2060. Expanding renewable energy in the four regions of the UK is seen as one major way in which such targets can be realised, since the government has indicated a reluctance to expand nuclear power generation. Nuclear power currently accounts for the biggest current share of electricity generation (44%) in Scotland: however, all of this capacity is planned to close by 2023. The primary policy instrument utilized by the Scottish Executive to promote the expansion of renewable energy sources is the Renewables Obligation (Scotland) (ROS). The ROS has combined a demand-push legal requirement for renewable power usage with a supply-pull financial incentive program to reward private industry for constructing and investing in new renewable energy generation projects (HMSO, 2002). The ROS compels electricity suppliers to source specific quantities of eligible renewable energy for sale to all customers (residential, commercial and industrial), or face financial penalties for the shortfall. The original minimum supply of renewable power by retailers, by quantity, was set at 3% for 3
2002-3, rose to 4.3% for 2003-4, and is expected to rise annually to 15.4% in 2015-2016. Since the ROS was implemented in April 2002, a significant increase has occurred in the number of renewable generation projects applying for planning consent in Scotland (BWEA, 2003).
3. Choice Experiments as a Means of Evaluating the Impacts of Renewables Policy. Renewable energy investments in Scotland are therefore expected to grow rapidly in the near future. These investments will produce a series of potential impacts on the environment and on the price of electricity. Environmental impacts include: •
landscape effects, such as through the construction of new wind farms on hillsides;
•
effects on wildlife, such as disruption of habitats and salmon migration; and
•
changes in air pollution, for example, air pollution from waste-to-energy plants, or reductions in air pollution by displacing fossil fuel plants..
Actual future environmental impacts (and changes in electricity prices through changes in generation costs) will depend on the exact investment mix which is undertaken (for instance, on the balance between on- and off-shore wind farms). Efficient social planning of which investment mix to promote should incorporate information on public preferences over these impacts. Environmental effects, price effects and indeed employment effects can be thought of as the attributes of a renewable energy strategy. The Choice Experiment (CE) method is a suitable as a means of estimating these public preferences over these attributes, since it is based on Lancaster’s “characteristics theory of value”. According to Lancaster, demand is defined over the characteristics of goods, rather than over goods themselves. In any CE exercise respondents are thus asked to choose between different bundles of goods described in terms of their characteristics (or attributes) and the levels these take. Offered choices defined in terms of these attributes, utility maximising individuals will choose the alternative that gives the highest level of utility: that is, individual
4
n will chose alternative i over some other option j if Uin > Ujn. In CE, this choice is represented using random utility theory (McFadden, 1973), whereby the conventional utility function U(.) is split into two parts: one deterministic part V(.) that contains factors observable by the analyst, and a random component ε (.), that represents determinants of individual choice that are not observable. In other words, the utility for individual i choosing alternative n is:
Uin = Vin + εin
(1)
with the probability that individual n will choose option i over any other option j belonging to some choice set C of:
Probin = Prob
(Vin + εin > Vjn + εjn)
∀j∈C
= Prob {(Vin - Vjn ) > (εjn - εin )}
(2a) (2b)
Since the random element of utility is by definition not observable, the analyst must make assumptions about the nature of the error component if he/she wishes to estimate the choice probability of (2). A typical assumption is that the error terms are independently and identically distributed with a Type 1 extreme-value (sometimes called Gumbell or Weibull) distribution. Under this hypothesis an explicit form of the probability of choice in equation (2) is satisfied by the Multi-Nomial Logit (MNL) model and is equal to:
Probin = exp(λVin) / ∑j exp (λVjn)
∀j∈C
(3)
where λ is a scale parameter which is inversely proportional to the standard deviation of the error terms, and Vin is the deterministic component of the utility function, assumed to be linear in parameters:
5
Vjn = ∑kβjk Xjk
(4)
where Xjk is the kth attribute value of the alternative j and βjk is the coefficient associated with the k'th attribute. The β parameters cannot in themselves be estimated, since they are not separable from the scale parameter. However, the marginal rate of substitution between any pair of attributes is obtainable, since the scale parameter cancels out, as is shown in (5) below:
MRS = - (λ.β attribute a / λ.β attribute b) =
- (β attribute a /β attribute b)
(5)
In designs where the cost or price of choosing an alternative has been included as an attribute, then (5) can be used to produce an estimate of the "implicit price" or “part-worth” P*a by replacing the denominator with the β estimate for this cost/price attribute:
P*a
= - ( β a / β price )
(6)
These implicit prices express the marginal WTP for a discrete change in an attribute level, and thus allow some understanding of the relative importance that respondents give to attributes within the design. Finally, compensating surplus welfare measures can be obtained for different environmental policy scenarios associated with multiple changes in attributes, using (7), where V0 is the calculated value for the deterministic part of U(.) under initial conditions, and V1 is deterministic utility evaluated under changed conditions. Again, the scale parameter cancels out in this expression (Bennett and Adamowicz, 2001), and Compensating Surplus, CS, is:
CS = - 1/β price (V0 - V1).
(7)
6
One important feature of the MNL model is that it assumes preferences are homogenous, since only a single parameter is estimated for each attribute included in V. However, we might well want to allow for tastes for a given attribute to vary over households. What is more, an individuals' preferences are likely to be unvarying over each choice she makes (for instance, over the four choice sets in our survey), even if preferences vary across individuals. The random parameters logit (RPL) model allows for such variation in preferences across households, and can adjusted to allow for error correlation across the choices made by each respondent (Meijer and Rouwendal, 2000; Revelt and Train, 1998; Train, 2003; Wedel and Kamakura, 2000). Denote U as the utility household n obtains from alternative i of a choice problem: Uni = µn′xni + eni
(7)
where xni is a vector of random variables, µn is a vector of coefficients and eni is the random component of utility. Subscript n in µn indicates that preferences for each random variable varies over households, which also affects the correlation between alternatives due to the common influence of µn over i. Consider now RPL choice probabilities such that µn varies in the population with density f(µn⎮Ω*), where Ω* is the true parameter vector for this distribution. Since µn is randomly distributed, we need to integrate the conditional logits evaluated over all values of µn to derive an “unconditional” logit probability: Qni(µn) = ∫ Pni(µn)f(µn⎮Ω*)dµn
(8)
where Pni(µn) = exp(µn′xni)/∑jexp(µn′xni) since the utility is linear in µn. The integral in this unconditional logit cannot in general be evaluated analytically. Nevertheless, one can resort to simulation to approximate the probabilities (Brownstone and Train, 1999; Train, 2003. In order to include the possibility of correlations over repeated choices by an individual household, introduce subscript t and describe the utility, conditional probability and unconditional probability respectively as follows: Unit = µn′xnit + enit Pnit(µn) = ∏t{exp(µn′xnit)/∑jexp(µn′xnit)}
7
(9) (10)
Qnit(µn) = ∫ Pnit(µn)f(µn⎮Ω*)dµn
(11)
These specifications explicitly incorporate the possibility of correlation between repeated choices, given the common influence of µn. It is important to note that the conditional choice probability is now the product of the formulas, evaluated each time. In order to consider a more specific source of taste variation, decompose the kth random coefficient, µkn, into the structural parameters:
µkn = γk + δk′wn + σkvkn
(12)
where γk is the constant, and wn is a vector of observed household-specific (or choicespecific) characteristics and thus δk is a vector of coefficients that give a household-specific mean. The random term vkn is an iid standard normal deviate: vkn ~ N(0,1), and σk represents the standard deviation of the marginal distribution of vkn. For the full vector of random coefficients, we can develop (12) as follows:
µn = γ + ∆wn + Λvn
(13)
where γ is a vector of constants, ∆ is a matrix of coefficients, wn denotes the householdspecific characteristics vector as above, and Λ is a vector such that we have the full covariance matrix Ω = ΛΛ′. Where the random coefficients are independent, Λ contains nonzero diagonal elements and zero below-diagonal elements. If the random coefficients are freely correlated, Λ contains non-zero, diagonal and below-diagonal elements for Ω. As such, the constant vector γ represents the population means of the random variables, ∆ produces heterogeneity in the means of the random variables depending on the household-specific characteristics, and Λ shows the standard deviations of the random variables , that is, the deviations of the tastes from the population means. As with the MNL model, the ratio between the coefficient for the price and the nonprice variables gives an implicit price for RPL model. However, the fact that RPL allows for a variety of assumptions about the distribution of coefficients often poses problems. Normal and lognormal distributions are often assumed for random coefficients. If the price and the
8
non-price random variables are assumed normal (or log-normal), then the implicit price follows the same distribution. If they assume different distributions, then the calculation is not so straightforward and the solution may even be undefined due to the non-linearity of the implicit price function.
4. Study Design and Implementation 4.1 Design of the Choice Experiment In any choice experiment, attributes must be chosen which meet a number of requirements. These are that they are relevant to the problem being analysed, realistic, capable of being understood by the sample population, and are applicable to policy analysis. Identifying the set of attributes and the levels these take is a key phase in choice experiment design. To this effect, focus groups were conducted with members from the general public (Dewar, 2003). These groups were used to identify the kinds of environmental impact of renewable energy which the general public were aware of, and thought important. These included: •
impacts on landscape – for example, from the construction of new wind farms;
•
impacts on wildlife, such as the effects of wind farms on birds, or of hydro schemes on salmon; and
•
pollution from waste-to-energy plants.
Focus group members also identified the employment effects of renewable investments as being very important as shaping their support or opposition to individual projects. Thus, although employment effects are not changes in the supply of public goods for which welfare measures can be constructed, they were included in the choice set design since they appeared to be an important determinant of attitudes. Finally, participants expressed concerns over the effects of increasing the proportion of electricity generated by renewables on the price paid by
9
households for electricity. This constitutes the price element of the design, in terms of equation (6). The attributes selected for the choice experiment design were thus impacts on the landscape, impacts on wildlife, impacts on air pollution, creation of long-term employment opportunities, and potential increases in electric prices. These are summarised in Table 1, along with the levels each could take. Note that these levels are, in some senses, qualitative and quite general. For instance, we refer to “slight harm” to wildlife population, and to “modest” impacts on landscape. This was due to the desire to represent the impacts of a wide range of renewable investments – across which effects would vary greatly; and the imprecision with which focus groups discussed landscape and wildlife effects. SPSS (Version 10.0) was used to create choice profiles using an orthogonal design, fractional factorial procedure. These choice profiles were then combined to make up the choice sets used in the experiment. Given the 5 attributes and 17 levels, 24 choice profiles were identified. These were alternated in the order in which they appeared as a choice set, and in terms of whether they were included as the first or second option. A zero additional cost option was included in each choice set as a status quo: this involved no further investments in renewables, investment in fossil fuel power instead, and no consequent increase in electricity prices. Figure 1 provides a representative choice set.
>
The survey instrument was developed using insights from the focus groups and from a small pilot survey. The questionnaire began by presenting the context of renewable energy development in Scotland. The national commitment by the United Kingdom to reduce production of global climate gases was explained. Survey participants were told that the survey was not concerned with any specific type of renewables technology, but with the impacts that could result from development of any renewable energy resource. The five attributes noted above were described, with examples being given to clarify each type of 10
impact. Four choice sets were then presented and the survey participant was requested to indicate their preferences between three options in each choice set The final page of the questionnaire was concerned with collecting standard socioeconomic information about the participant. Information was requested about residential location of household, number of children, employment in the energy sector, membership in a conservation group, age, household income, education attainment, and amount of last electric bill. 4.2 Sample Selection The target population for this survey was the Scottish general public. Our sample population was randomly drawn from a sampling frame of a list of registered voters in eight council districts of Scotland. Some 547 addresses were then mailed survey packages with a cover letter during the first week of September 2003. As an incentive to participate a £20 prize draw was offered. Three weeks later a follow-up postcard was mailed to encourage the completion and return of the survey. By end-October, 219 households had returned surveys, a 43% response rate after un-deliverables are considered. Any mail survey runs a risk of selection bias, in that those more interested in the environmental issue of concern, or perhaps those with higher levels of education and environmental awareness, are more likely than average to return their surveys. Comparing socio-economic information collected on the respondents used in the choice experiment against the statistical profile of the Scottish population is one test for such a bias: the null hypothesis that the sample characteristics are equal to the population characteristics must be rejected for bias to be suspected. In our sample, respondent’s income and location of residence are statistically different from the national distribution at 10% level. Our sample is thus lower income than the national average, and more rural. These two descriptors are in fact correlated with each other.
5. Model Estimation 5.1 Homogeneous Preferences: Multi-Nomial Logit 11
NLOGIT 3.0/LIMDEP 8.0 was used to estimate an MNL model. Environmental attributes were effect coded (Louviere, Henscher and Swait, 2000). Effect coding means that at least one level of each attribute is not included as a variable: thus a 3-level attribute generates two variables. The attributes levels chosen for exclusion were those hypothesised to have the most negative effect on environmental amenities. The estimated coefficients for each of the included levels thus indicate the value respondents placed on a change from the lowest valued (omitted) level. The omitted levels were: High Landscape Impact, Slight Wildlife Harm, and Slight Increase in Air Pollution. Results for all respondents from the MNL model are shown in Table 2. The “attributes-only” model shows results when only the choice experiment attributes are included in the regression. All attribute coefficients have the expected signs. The signs of all but the price attribute are positive as expected, since all are coded to show an increase in environmental quality which should lead to increased utility. Price is negatively signed: respondents prefer cheaper electricity, everything else equal. All of the environmental attributes are significant determinants of choices at the 95% level of confidence. However, employment creation is not a significant determinant of choice. Several socio-economic variables were tested for inclusion in an "expanded" version of the MNL model. Three covariates were selected to be in the "expanded" model (Table 2): these either show statistical significance, or are included on theoretical desirability grounds. Education and age are in the former class, while income is the latter case. A likelihood ratio test was used to compare the "attributes only" and "expanded" models, and rejected the null hypothesis that the parameter values of the two models are equal at the 95% significance level. The adjusted McFadden Pseudo-R2 is also improved with the addition of the socioeconomic covariates. Louviere et al (2000) state that a McFadden statistic in the 0.20 to 0.40 range, is comparable to an ordinary least squares (OLS) adjusted-R2 of 0.70 to 0.90 in range. Therefore the expanded model with covariates is deemed the superior model, and implicit prices from this are used in the following discussion. >
12
Implicit prices (equation 6) are interpreted as the willingness-to-pay through an increase in electricity charges per annum per household, for a level change in any of the attributes. Estimates are shown in Table 3. With regard to the level of landscape impacts, moderate and low landscape impacts were not statistically significant compared with a high impact. Respondents thus only seem WTP to reduce high landscape impacts. Scope effects are present for wildlife impacts, in that respondents are willing to pay more for an increase in wildlife habitat than for preventing any decline. Wildlife improvements are valued more highly than avoiding high landscape impacts. Avoiding increases in air pollution, through an expansion of biomass plants for instance, has the highest WTP estimate across the environmental attributes. Interestingly, this ordering of implicit prices matches responses to a separate survey question whereby respondents were asked to indicate which single attribute was most important to them. The ordering of the attributes by votes from respondents was: air pollution, wildlife, electricity price, landscape, and employment. This shows consistency with the ordering of WTP estimates in Table 3.
>
5.2 Allowing for Heterogeneous Preferences One important factor that may determine one's preferences over the impacts of renewable energy projects is where one lives, in particular whether one lives in the countryside or not. A way of testing this in our survey is to examine whether there is a statistical difference between rural and urban estimated MNL coefficients and implicit prices. To do this, the sample was partitioned according to place of residence as disclosed in the questionnaire. The sample population was thus segregated into two groups, those located in villages or the countryside and those who reside in towns and cities. Separate MNL models were then run for each group (Table 4). A likelihood ratio test rejected the null hypothesis that the segregated subsets were equal at the 5% level. Moderate landscape impacts now register as significant in the rural model, as do jobs. Jobs remain insignificant in the urban sample, but 13
become strongly significant in the rural model: this is perhaps unsurprising given most peoples' likely expectations about where jobs would be created.
>
Another reason why attitudes towards renewable energy investments might vary across people is their income: either because environmental concern is a "luxury" (Hokby and Soderqvist, 2003), or because rising energy prices hit poorer households disproportionately hard. To test this hypothesis, the sample was split by annual household income level into two sub-samples: low income (£16,000 or less per year), and higher income (greater than £16,000 per year). However, a log-likelihood ratio test failed to reject the null hypothesis that the two subsets were equivalent: there are no significant differences in preferences therefore between these two income groups. Splitting the sample by income or place of residence is an exogenously-imposed means of allowing, in a rather limited way, for preference heterogeneity. A rather more general approach, which allows for endogenous identification, is to use the Random Parameters Logit (RPL) model, as explained in Section 3. LIMDEP was therefore used to fit two RPL models to the data (Table 5). These models incorporate only the Alternative Specific Constants and the choice attributes. The first RPL model rules out the possibility of correlation between the repeated choices of each respondent (non-panel RPL), whereas the second allows for this (panel RPL). Both RPL models assume normal distributions for the variables other than the ASCs and the price. The ASCs are fixed, whilst preferences towards the price attribute are assumed to be lognormal (since every household is assumed to place a negative value on an increased price).
> The most interesting result that emerges from Table 5 is that few practical differences emerge between the two RPL models and the simple MNL model. The population means of 14
all attribute coefficients have the a priori expected signs. All of the environmental attributes turn out to be significant determinants of utility. The MNL and the panel models have very similar levels of significance for the mean effects, with the non-panel model showing slightly lower significance levels, especially for the impacts of renewable energy projects on wildlife. The creation of jobs is insignificant for all models except the panel RPL. All models also show significant ASCs, indicating that factors outside the models tend to make people prefer renewable energy projects to alternative, non-renewable investments. It is important to note that all but one standard deviation terms for the RPL coefficients are insignificant. This result suggests that households may in fact have rather similar tastes for the potential impacts of renewable energy projects on the environment, that is, that preferences are rather homogeneous. In particular, the standard deviations for the panel model are very small. This indicates that, even where households are assumed to rely on constant tastes over repeated choices, taste variation may be too small to have meaningful impacts on choice probabilities. Lastly, none of the insignificant mean effects is matched with a significant standard deviation effect. This result suggests that the insignificant mean effects may not be attributable to the possibility that significantly different tastes for a given attribute cancel out across the sample. All models have practically the same level of fit – the adjusted Pseudo-R2 values are about 0.30 in all cases. As can be seen from Table 5, RPL tends to give consistently larger coefficients than MNL. The stochastic portion of utility for RPL is decomposed into the taste variation (σkvkn) and the error term (e) whilst the stochastic portion for MNL is composed of the error term (e) alone. This larger error term in MNL than in RPL leads to smaller coefficients for MNL than for RPL because of normalisation (Revelt and Train, 1998). As such, an interesting issue arises: a disproportional increase of non-price coefficients to the price coefficient would lead to a disparity in the implicit prices between MNL and RPL. We estimated several other RPL models for comparison purposes. However, allowing for random ASCs, a normally-distributed price, and a fixed price all gave practically the same results as the above. We also examined the potential effects of socioeconomic variables 15
included in wn - income, rural or urban residence, age, and education – on the heterogeneity of the means of the random variables. Almost all of the results are insignificant, except in three cases. Older people are expected to place higher values on reducing air pollution and lower values on the number of jobs created. More educated people are likely to feel lower disutility from an increased electricity bill. Note however that these socio-economic variables are only significant at the 10% level. This implies that, once we allow for the possibility of taste variations through the RPL technique, the influence of socio-economic terms seen in the MNL model reduces. This is to be expected. Table 6 calculates the implicit prices using the models reported in Table 5. In the case of the RPL models, the means of the attribute coefficients are divided by the mean of the price coefficient, as with the MNL models. Since the price coefficient is lognormal, its mean coefficient was evaluated at exp(m + s/2), where m and s respectively denote the mean and variance of log of the price coefficient. The estimates for the implicit prices are somewhat different between the MNL and RPL models. On the whole, implicit prices for the non-panel RPL are bigger than for the MNL, whereas the panel RPL generates both bigger and smaller implicit prices than the MNL. However, for each attribute, the confidence intervals for the estimates overlap considerably across the models. Overall, then, we find little change in model fit, coefficient estimates or implicit prices in moving from the MNL model to either of the RPL models.
>
6. Welfare Analysis for Alternative Investment Scenarios One of the strengths of choice experiments is that estimated coefficients of the attributes maybe used to estimate the economic value of different ways in which the attributes can be combined. To determine the change in compensating surplus from possible alternative projects, the expression given in equation 7 is evaluated (using the MNL model), as a means of comparing welfare in a base-case situation and in some alternative situation. Here, the base 16
case was chosen to be no increase in renewable power, but an expansion of fossil-fuel power instead. Four alternative investment scenarios in renewables were used, labelled A, B, C and D:
Scenario A: a large (200 MW) Offshore Windmill Farm, with 100 turbines each at 80 meters nacelle hub height, 6-10 kilometres from shore. These figures are based on projects currently under planning consideration in the UK (BWEA, 2003). This project is assumed to have no landscape effect, no harmful effect on wildlife, and to reduce air pollution relative to the base case. Scenario B: a large scale (160 MW) Onshore Windmill Farm, with 80 turbines each at 80 meters nacelle hub height. This project description is an average of the permitted large onshore windfarms that are entering the construction phase during 2003-2005 in Scotland. It is assumed to have a high landscape impact, to cause slight harm to wildlife (eg to migratory geese), and to reduce air pollution relative to the base case. Scenario C: a small scale (50 MW) Windmill Farm with 30 turbines each at 60 meters nacelle hub height. This is assumed to have a moderate landscape impact, to cause no harm to wildlife, and to reduce air pollution relative to the base case. Scenario D: a Biomass Power Plant with a capacity of 25MW, and an emissions stack height up to 40 meters, fuelled by energy crops. This project description is extrapolated from a biomass power plant presently going through the planning process, with a fuel source of coppiced willow grown in the surrounding agricultural region. It is assumed to have a moderate landscape effect due to its construction, to enhance wildlife habitat through the growing of coppiced willow, but to increase air pollution relative to any of the three wind power projects. Table 7 gives results. Note that we omit employment effects from these welfare change calculations, since changes in employment are not comparable with changes in environmental public goods in utility terms. The values for welfare change shown can be interpreted as the amount that households are willing-to-pay on an annual basis to have 17
renewable energy projects with the indicated attribute levels, rather than the base case expansion of fossil fuel power. Scottish households place the greatest value on offshore wind farms, with the next most valued type of energy project being a small scale on-shore windfarm. The major determinant of compensating surplus for both of these projects is the avoidance of air pollution from fossil-fuel combustion, and the fact that high landscape effects are avoided. The biomass power plant has the lowest compensating surplus, due to its air pollution effects, despite the small improvement in wildlife habitat which results. Finally, willingness to pay for the large scale windfarm is negative, since the dis-utility impacts (principally on landscape) outweigh the positive impacts on air pollution. We also report in Table 7 the welfare cost per MW of installed power when the choice experiment results are aggregated to the Scottish population. As may be seen, the large off-shore wind farm option emerges as a clear “winner” in terms of environmental benefits relative to generating capacity installed, and that the relative case for the biomass plant is improved since its power capacity is higher than the small on-shore wind farm.
>
8. Conclusions Renewable energy offers a partial solution to the problem of reducing greenhouse gas emissions whilst meeting future energy needs. Yet different renewable energy projects can have varying external costs in terms of impacts on the landscape, on wildlife and on air pollution. In addition, strategies vary in their likely impacts on jobs and electricity prices. The choice experiment method used in this paper enables these effects to be jointly evaluated. This enables conclusions to be drawn about the net social benefits of different renewables investment strategies. Reviewing results from the MNL model, we found a substantial sensitivity to the creation of projects that will have a high impact on landscapes. Conversely, there seems to be 18
no sensitivity, or at least no positive mean willingness-to-pay, to reduce landscape impacts if the projects are designed to have lower levels of landscape effects. Results show that increases in wildlife, such as might be generated by pursuing biomass options, attract a positive economic value. Conversely, avoiding air pollution from renewable energy investments was also valued by our respondents. This would add to the case against burning biomass for power. Tacking steps to allow for preference heterogeneity gives mixed results. When separate MNL models are estimated for rural and urban respondents, we find that these models produce significantly-different estimates of preferences, using a Likelihood Ratio test. This is not the case when the sample is split according to income group. However, when a Random Parameters Logit model which allows endogenously for evidence of preference heterogeneity to emerge is used, little such evidence is found, since the standard deviation terms on the environmental impact attributes are, on the whole, insignificant. We found little effect on either explanatory power, coefficient size or implicit prices in moving away from the MNL model to a RPL set-up with or without correlation across choices. In our case, it thus seems that there is little good reason to move away from the simple MNL framework as a model of preferences over renewable energy alternatives (Louviere, Henscher and Swait, 2000). Finally, we note that the comparison of welfare costs and benefits given in Table 7 makes use of the installed capacity of the four renewable power options considered. Actual outputs from these will obviously be (considerably) less this, so that an analysis of welfare effects per MWhr delivered might be of interest in a future study.
Acknowledgement: We thank the Scottish Economic Policy Network for funding the research on which this paper is based.
19
References Alvarez Farizo B and Hanley N, (2002) Using conjoint analysis to quantify public preferences over the environmental impacts of wind farms. Energy Policy, 30 (2), 107-116. Dewar, M. (2003) Report for the University of Glasgow Renewable Energy Project: Synopsis of Focus Group Responses. New Lanark: Focus Ecosse. Dti (2003) Energy White Paper “Our energy future – creating a low carbon economy”, London, HMSO. Hanley N and Nevin C (1999) “Appraising renewable energy developments in remote communities: the case of the North Assynt Estate, Scotland.” Energy Policy, 27, 527547. Hanley N,, Bergmann A and Wright R (2003) Valuing the Attributes of Renewable Energy Investments in Scotland. Report to the Scottish Economic Policy Network, available at www.scotecon.net. HMSO (2001) The Renewables Obligation (Scotland) Order 2002. Scottish Statutory Instrument 2002 No. 163. Edinburgh, HMSO. Hokby S. and Soderqvist T (2003) "Elasticities of demand and willingness to pay for environmental services in Sweden" Environmental and Resource Economics, 26 (3), 361-383. Lancaster, K. (1966) ‘A New Approach to Consumer Theory’, Journal of Political Economy, 74, 132-157. Louviere, J.J., D. Hensher and J. Swait (2000): Stated Choice Methods: Analysis and Application, Cambridge: University Press. McFadden, D. (1973). “Conditional Logit Analysis of Qualitative Choice Behaviour.” In: P. Zarembka (ed.), Frontiers in Econometrics. New York: Academic Press. McFadden, D. and K. Train (2000), “Mixed MNL Models for Discrete Responses,” Journal of Applied Econometrics, Vol. 15, pp. 447~470. 20
Meijer, E. and J. Rouwendal (2000), “Measuring Welfare Effects in Models with Random Coefficients,” SOM-theme F: Interactions between Consumer and Firms. AKF, Copenhagen. Owen A. D. (2004) “The transition to renewable energy” in A. Owen and N. Hanley (eds.) The Economics of Climate Change. London: Routledge. Revelt, D. and K. Train (1998), “Mixed Logit with Repeated Choices: Households’ Choices of Appliance Efficiency Level,” Review of Economics and Statistics, Vol. 80, No. 4, pp. 647~657. Scottish Executive (2002a) Key Scottish Environmental Statistics, 2002. A Scottish Executive National Statistics Publication, Edinburgh. Scottish Executive (2002b) Securing a Renewable Future: Scotland’s Renewable Energy HMSO, Edinburgh. Train, K. (2003), Discrete Choice Models with Simulation, Cambridge, Massachusetts: Cambridge University Press. De Vries HJ, Roos CJ, Beurskens LW, Kooijman, AL and Uyterlinde MA (2003) Renewable electricity policies in Europe. Amsterdam: Energy Research Centre of the Netherlands. Wedel, M. and W. Kamakura (2000), Market Segmentation: Conceptual and Methodological Foundations, 2nd Ed., Boston, MA: Kluwer Academic Publishers.
21
Table 1.Attributes, descriptions and levels
Attribute
Description
Levels
Landscape Impact
The visual impact of a project is dependent on
None, Low
a combination of both the size and location.
Moderate, High
Change in habitat can influence the amount and
Slight Improvement,
diversity of species living around a project.
No impact, Slight
Wildlife Impact
Harm
Air Pollution
Many types of renewable energy projects create no
None, Slight increase
additional air pollution, but some projects do burn non-fossil fuels. These projects produce a very small amount of pollution when compared to electricity generated from coal or natural gas.
Jobs
All renewable energy projects will create new
1-3, 8-12, 20-25
local long-term employment to operate and maintain the projects. Temporary employment increases during the construction phase are not being considered.
Price
Annual increase in household electric bill resulting from expansion of renewable energy projects. An average household pays £270 a year (£68 per quarter) for electricity
Alternate specific constants ASC-A
Takes value of 1 for Plan A, 0 otherwise.
ASC-B
Takes value of 1 for Plan B, 0 otherwise.
22
£0, £7, £16, £29, £45
Figure 1.
Example choice set
option example Plan A
Plan B
Neither
No increase in
LANDSCAPE HIGH
NONE
SLIGHT HARM
SLIGHT HARM
visual impact caused by
renewable energy
location and/or size
WILDLIFE health of habitat
Alternative
AIR POLLUTION
NONE
NONE
climate change programs used
EMPLOYMENT
8-12 JOBS
1-3 JOBS
new jobs in local community
£
North Sea gas
PRICE OF ELECTRICITY
£16
£7
fired power
per year
per year
stations instead
A
B
I would not want
additional rates per year
YOUR CHOICE:
either A or B
(please tick one only)
23
Table 2: Multi-nomial model results Coefficient estimate:
Coefficient estimate:
Attributes-only model
Expanded model
Landscape effects: moderate
0.20
0.29
Landscape effects: low
0.16
0.15
Landscape effects: none
0.39*
0.42*
Wildlife effects: no impacts
0.27*
0.22**
Wildlife effects:
0.50*
0.63*
Air pollution: no additional
0.71*
0.74*
Employment (number of jobs
0.01
0.02
-0.05*
-0.05*
Income * ASC A
-
-0.01
Income * ASC B
-
-0.01
Education * ASC A
-
0.99*
Education * ASC B
-
0.85*
Age* ASC A
-
1.06*
ASC A
2.96*
2.80*
ASC B
2.80*
2.73*
improvement in habitat
created) Rise in electricity prices
Number of observations
836
744
"pseudo R-squared"
0.29
0.31
** significant at 99% * significant at 95%
24
Table 3: Implicit Prices from the MNL model with co-variates, full sample (all WTP figures are £ per household per year)
Landscape Impact
Households are WTP £8.10 ( 95% confidence interval £4.30£11.90) to avoid a high impact landscape change relative to no landscape impact.
Wildlife Impact
WTP of £4.24 (£0.03-£8.51) to change a slight increase in harm to wildlife from renewable projects to a level that has no harm. However, households would be WTP £11.98 (£8.30-£15.66) to change a slight increase in harm to wildlife from renewable projects to a situation where wildlife is improved from the current level.
Air Pollution Impact
Households are WTP £14.13 (£10.45-£17.81) to have renewable energy projects that result in no increase in air pollution, compared to a programme which results in a slight increase in pollution.
25
Table 4. Implicit Prices for changes in attributes comparing rural and urban respondents (95% confidence intervals in brackets).
Landscape effects: from high to moderate Landscape effects: from high
Willingness to Pay value,
Willingness to Pay value,
rural people (£/hsld/yr)
urban people (£/hsld/yr)
12.15
N/S
[1.90 - 22.40] N/S
N/S
N/S
8.73
to low Landscape effects: from high to none Wildlife effects: from
[4.01-13.45] N/S
N/S
15.23
7.62
[9.04-21.49]
[2.87-12.36]
19.09
11.77
[11.77-26.39]
[7.70-15.85]
1.08
N/S
harmed to neutral Wildlife effects: from harmed to improved Air pollution: no additional air pollution rather than a small increase Employment (number of jobs created)
[0.20-1.95]
note: N/S means the effect was not significant in the MNL model, so that willingness to pay for this attribute change may well be zero
26
Table 5. Comparing MNL and RPL results
MNL Mean coefficient Moderate Landscape impact Low Landscape impact No Landscape impact
No Wildlife impact
0.21 (0.14) 0.15 (0.19) 0.39* (0.10)
Non-panel RPL Std.dev. Mean of of Coefficient coefficien t 0.27 0.19 (0.25) (1.67) 0.25 0.62 (0.33) (1.37) 0.63** 1.44** (0.28) (0.71)
Panel RPL Std.dev. Mean of of coefficient Coefficien t 0.29 0.0002 (0.20) (0.0005) 0.01 0.0003 (0.26) (0.0008) 0.34* 0.0002 (0.13) (0.0004)
0.29* (0.10) 0.49* (0.12)
0.48*** (0.27) 0.71** (0.30)
0.34 (0.91) 0.70 (0.92)
0.30** (0.14) 0.54* (0.17)
0.00004 (0.0003) 0.0002 (0.0005)
No additional air pollution
0.71* (0.06)
1.01* (0.34)
0.46 (0.60)
0.77* (0.08)
0.0001 (0.0002)
Employment
0.01 (0.01)
0.01 (0.02)
0.04 (0.03)
0.22*** (0.12)
0.00003 (0.00003)
Improved Wildlife
Price (negative) ASCA ASCB Log-likelihood (const.) Pseudo-R2 (adjusted)
0.05* (0.006) 2.89* (0.47) 2.75* (0.47) -505.54 0.2904
-2.72* 0.12 (0.34) (0.52) 4.15* (1.43) 3.97* (1.38) -501.29 0.2929
-2.94* 0.0004 (0.18) (0.0007) 3.23* (0.50) 3.08* (0.49) -499.74 0.2951
Notes: * Significant at 1% level, ** Significant at 5% level, *** Significant at 10% level. The figures for Price in RPL models represent ln(original coefficient). Standard errors in parentheses.
27
Table 6. Comparing implicit prices from the MNL and RPL models (values are £/household/year)
Moderate Landscape Impact Low Landscape Impact No Landscape Impact No Wildlife Impact Improved Wildlife No Air pollution impact Employment
MNL
Non-panel RPL
Panel RPL
Mean Std.err. (95% CI)
Mean Std.err. (95% CI)
Mean Std.err. (95% CI)
4.24 3.03 (-1.69 – 10.18) 3.02 3.63 (-4.10 – 10.13) 7.96* 1.97 (4.09 – 11.83)
4.03 3.80 (-3.42 – 11.48) 3.73 4.49 (-5.08 – 12.54) 9.51* 2.93 (3.76 – 15.26)
5.49 4.03 (-2.41 – 13.39) 0.21 4.85 (-9.30 – 9.73) 6.43* 2.19 (2.14 – 10.72)
5.88* 2.23 (1.50 – 10.25) 9.98* 1.91 (6.23 – 13.73)
7.21** 3.41 (0.52 – 13.90) 10.66* 2.23 (6.28 – 15.04)
5.76***3.14 (-0.39 – 11.90) 10.30* 2.41 (5.59 – 15.01)
14.45* 1.91 (10.70 – 18.19)
15.24* 2.89 (9.57 – 20.91)
14.62* 2.79 (9.14 – 20.10)
0.20 0.23 (-0.24 – 0.65)
0.17 0.30 (-0.42 – 0.75)
0.43*** 0.25 (-0.06 – 0.92)
Notes: * Significant at 1% level, ** Significant at 5% level, *** Significant at 10% level. The delta method was used to approximate the standard errors.
28
Table 7: Welfare Changes for Alternative Investment Strategies
Scenario:
Base Case A Fossil Fuel power Large Offshore station Wind farm expansion Attribute Levels: Effects on Low None Landscape Effects on None None Wildlife Change in Air Increase None Pollution Welfare Change (£/hsld/yr.) Welfare change relative to MW of power capacity, for Scottish households*
B
C
D
Large Onshore Small Onshore Biomass Wind farm Wind farm Power Plant High
Moderate
Slight harm
None
Moderate Slight improvement
None
None
Increase
+£33.00
-£10.60
+£29.20
+£5.40
+£2.88
-£7.18
+£0.81
+£2.20
* assumes 2.1 million households. Calculated as MW capacity for each scenario divided by £m of welfare gain/loss.
29
