Published Version

Testing Choice Theory Using Discrete Choice Experiments in Swiss Energy Policy

Matteo Mattmann ii

Promoter: Prof. Dr. Roy Brouwer Co-Promoter: Dr. Ivana Logar

Thesis committee: Prof. Dr. Wouter Botzen Prof. Dr. Michael Getzner Dr. Jürgen Meyerhoff Prof. Dr. Ståle Navrud Prof. Dr. Rolf Wüstenhagen

Cover Design: Åsa Frölander

ISBN: 978-3-906327-95-2 iii

VRIJE UNIVERSITEIT

Testing Choice Theory Using Discrete Choice Experiments in Swiss Energy Policy

ACADEMISCH PROEFSCHRIFT

ter verkrijging van de graad Doctor aan de Vrije Universiteit Amsterdam, op gezag van de rector magniﬁcus prof.dr. V. Subramaniam, in het openbaar te verdedigen ten overstaan van de promotiecommissie van de Faculteit der Bètawetenschappen op dinsdag 17 oktober 2017 om 11.45 uur in de aula van de universiteit, De Boelelaan 1105

door Matteo Mattmann geboren te Zürich, Zwitserland iv promotor: prof.dr. Roy Brouwer copromotor: dr. Ivana Logar

This PhD thesis was funded by the Swiss Federal Institute of Aquatic Science and Technology (Eawag), and is part of the Competence Center for Research in Energy, Society, and Transition (SCCER CREST). v

“Durchaus studiert, mit heißem Bemühn. Da steh ich nun, ich armer Tor! Und bin so klug als wie zuvor;”

J.W. von Goethe in Faust

vii

Summary

Testing Choice Theory Using Discrete Choice Experiments in Swiss Energy Policy

The "Swiss Energy Strategy 2050" proposes to phase-out nuclear power generation and expand renewable sources of energy. Hydropower, an established source of energy in Switzerland, is expected to be one of the key renewables that will be further expanded. In this policy context, this PhD thesis aims to test axioms and assumptions underlying microeconomic choice theory by applying discrete choice experiments (DCE). A DCE is conducted among a representative sample of Swiss respondents and elicits their preferences for an expansion of hydropower. This dissertation contributes to the existing literature by examining how public preferences for expanding hydropower production are linked to public perception of (avoiding) nuclear risks. To this end, hydropower as well as nuclear risks are included in the DCE. This thesis begins with a quantitative meta-analysis of the existing stated preference literature that estimates the non-market values of hydropower externalities. The results of the meta-analysis are used as inputs in designing the DCE. The results of the meta-analysis suggest that deteriorations in vegetation, landscape, and wildlife are valued negatively, while there is only limited evidence for a significant positive willingness-to-pay (WTP) for mitigating these negative externalities. The avoidance of greenhouse gas emissions proves to exert a significant positive influence on welfare estimates, but no significant impacts on aesthetic and recreational amenities can be detected. The meta-analysis also reveals that no stated preference studies so far have considered the link between preferences for renewable sources of energy and nuclear risks. The data obtained from the DCE are used to answer this dissertation’s main research questions. These focus on the standard choice-theory assumptions of viii certain and known preferences and the axioms of continuity and monotonicity. Furthermore, the role of multiple reference points in the framework of prospect theory is investigated. More specifically, the common and idiosyncratic determinants of choice certainty, consistency, and monotonicity are investigated. In contrast to the existing literature, these three concepts are analyzed simultaneously based on the same sample of respondents. The results show that there are significant differences between the choice behavior of certain and uncertain respondents as well as between consistent and inconsistent respondents. Moreover, gender and choice- task complexity prove to be common predictors of choice certainty, consistency, and monotonicity. This thesis also investigates the standard economic axiom of continuous preferences in the context of attribute-non-attendance (ANA). A novel methodology to assess ANA is presented based on the monitoring of the respondents’ visual information acquisition process using mouse-tracking. No significant model improvement is found when including such a visual measure of ANA compared with the standard approach based on stated ANA information. Nevertheless, choice models based on visual ANA result in a slight improvement over choice models that do not take ANA into account and over choice models that use inferred ANA information. Finally, the dependence of preferences on (multiple) reference points, a key assumption in prospect theory, is tested. Non-status quo related reference points, associated with comparative risks shown on risk ladders, are expected to affect parameter estimates and welfare measures for a change in hydropower and nuclear risk. The study confirms the importance of multiple reference points, and shows that, besides the status quo, these other reference points also influence respondents’ choices and welfare measures in DCEs. The results of this thesis support the need for a holistic view on energy policy accounting for the direct and indirect externalities of alternative energy sources in both research and policy. ix

Acknowledgements

First and foremost I would like to thank my supervisors Ivana Logar and Roy Brouwer. It has been a very supportive, friendly, and uncomplicated collabora- tion with both of you. Your comments and ideas often perfectly complemented each other. Ivana, a special word of thanks to you for the infinite amount of time you have invested in giving me very valuable, extensive, and precise feedback on work in progress at various stages. Roy, special thanks to you for your enthu- siasm and for keeping the big picture in sight at times when no light at the end of the tunnel seemed visible. Thank you also for putting me up on various visits to Amsterdam and Waterloo in Canada, and for connecting me with the people in your respective teams. Many thanks also belong to the members of the reading committee for the review of this dissertation: Wouter Botzen, Michael Getzner, Jürgen Meyerhoff, Ståle Navrud, and Rolf Wüstenhagen. Special thanks are due to Mehmet Kut- luay, my connection to the Vrije Universiteit in Amsterdam and the team at IVM, for the motivating and valuable exchange. Thank you Noémie Neverre for the excellent French translation of my survey. I would also like to thank conference attendees in Nancy, Zurich, Cork, and Athens, as well as the organizers and members of the Competence Center for Research in Energy, Society and Transi- tion (SCCER CREST), of which this thesis is part of. This research project was funded by the Swiss Federal Institute of Aquatic Science and Technology (Eawag), and I would like to thank Eawag. A number of great people made the institute an interesting and enjoyable place to work. First of all, there are my team- and office-mates Paola and Markus. Thank you Paola for the humor you brought to our office and for your wise advice regarding finishing a PhD. Thank you Markus for the interesting and useful discussions. Thank you dear other members of ESS: Alex, Alice, Bernhard, Caroline, Fridolin, Jasmine, Mara, Mario, Mika, Mirella, Pauline, Philipp, Simon J., Simon M., and Ulrike. Last but not least, thank you Maja for all the food supply. Thank you.

Contents

Summary vii

Acknowledgements ix

1 Introduction 1 1.1 Background ...... 1 1.2 Main objective, hypotheses, and research questions ...... 3 1.3 Data collection and econometric analysis ...... 6 1.4 PhD thesis outline ...... 8

2 Hydropower Externalities: A Meta-Analysis 11 2.1 Introduction ...... 11 2.2 Study selection and characteristics ...... 15 2.3 Meta-model ...... 19 2.3.1 Heterogeneity, heteroskedasticity, and non-independence 19 2.3.2 The meta-regression models ...... 20 2.4 Selection and deﬁnition of variables ...... 22 2.5 Results ...... 27 2.5.1 Descriptive statistics ...... 27 2.5.2 Meta-regression results ...... 29 2.5.3 Cross-validation ...... 33 2.6 Conclusions and discussions ...... 35 2.A Studies included in the meta-analysis ...... 38

3 Choice Certainty, Consistency, and Monotonicity 41 3.1 Introduction ...... 41 3.2 Choice certainty, consistency, and monotonicity in DCEs ..... 44 3.3 Econometric Models ...... 48 xii

3.4 Case-study description ...... 50 3.4.1 Discrete choice experiment ...... 50 3.4.2 Elicitation of choice certainty, consistency, and monotonicity 52 3.4.3 Sampling procedure and choice experiment design .... 53 3.5 Results ...... 55 3.5.1 Descriptive results ...... 55 3.5.2 Swait-Louviere test results ...... 57 3.5.3 Logit model results ...... 62 3.6 Discussion and conclusions ...... 66

4 Attribute non-Attendance in Discrete Choice Experiments 69 4.1 Introduction ...... 69 4.2 Attribute non-attendance ...... 71 4.3 Case-study description and experimental design ...... 77 4.4 Elicitation of stated and visual ANA ...... 79 4.5 Econometric models ...... 81 4.6 Results ...... 84 4.6.1 Descriptive statistics ...... 84 4.6.2 Stated ANA models ...... 86 4.6.3 Visual ANA models ...... 89 4.6.4 Inferred ANA models ...... 91 4.7 Discussion and conclusions ...... 94 4.A Example choice task ...... 97 4.B Lookup frequency and duration ...... 98 4.C AIC and BIC ...... 99 4.D Baseline ECLC model ...... 100

5 Reference Points for the Valuation of Risk Changes 101 5.1 Introduction ...... 101 5.2 Theoretical framework and hypotheses ...... 104 5.3 Case-study description ...... 108 5.3.1 Choice experiment ...... 108 5.3.2 Risk ladders ...... 109 5.3.3 Covariates ...... 112 5.3.4 Design generation and data collection ...... 113 xiii

5.4 Econometric models and testing procedures ...... 115 5.5 Results ...... 116 5.5.1 Descriptive statistics ...... 116 5.5.2 Choice model results ...... 119 5.5.3 Hypotheses test results ...... 123 5.6 Discussion and conclusions ...... 124 5.A Poe-test ...... 127

6 Conclusions 129 6.1 Summary of the main ﬁndings ...... 129 6.2 Directions for future research ...... 132 6.3 Policy recommendations ...... 135

A Choice Experiment Survey 137 A.1 Welcome ...... 137 A.2 Your electricity consumption ...... 137 A.3 Hydropower and nuclear power ...... 139 A.4 Your opinion about dam breaches ...... 139 A.5 Your opinion about nuclear accidents ...... 140 A.6 Risk graph ...... 141 A.7 Expansion of hydropower ...... 143 A.8 Your opinion in a public vote ...... 144 A.9 How hydropower can be expanded (I/II) ...... 144 A.10 How hydropower can be expanded (II/II) ...... 145 A.11 Example situation ...... 147 A.12 Eight hypothetical decision situations ...... 147 A.13 Decision 1 ...... 148 A.14 Decision 2 ...... 149 A.15 Decision 3 ...... 150 A.16 Decision 4 ...... 151 A.17 Decision 5 ...... 152 A.18 Decision 6 ...... 153 A.19 Decision 7 ...... 154 A.20 Decision 8 ...... 155 A.21 Background of your decisions ...... 156 xiv

A.22 Your leisure time ...... 157 A.23 Trust and attitude ...... 158 A.24 About your person (I/II) ...... 159 A.25 About your person (II/II) ...... 160 A.26 On this survey ...... 162

Bibliography 163 xv

List of Figures

2.1 Cross-validation histograms ...... 34

3.1 Choice task example ...... 52 3.2 Stated choice certainty ...... 56 3.3 Choice consistency ...... 57

4.1 Example choice task of the mouse-tracking setup ...... 80 4.2 Descriptive statistics for mean stated ANA ...... 85 4.3 Example choice task ...... 97 4.4 Average lookup frequency over choice tasks ...... 98 4.5 Average lookup duration over choice tasks ...... 98 4.6 AIC and BIC ...... 99

5.1 Schematic illustration of two different risk ladders ...... 105 5.2 Expected changes in utility associated with reference points ... 107 5.3 Choice task example ...... 110 5.4 Risk ladders shown to two samples ...... 111 5.5 Descriptive statistics of risk attitudes and perceptions ...... 117

xvii

List of Tables

2.1 Studies collected in the selection and search procedure ...... 17 2.2 Explanatory variables included in the meta-analysis ...... 23 2.3 Cross-tabulation of mean values of hydropower externalities ... 28 2.4 Meta-analysis regression models ...... 32

3.1 Studies that regress stated choice certainty on its determinants .. 45 3.2 Studies that regress choice consistency on its determinants .... 46 3.3 Attribute and attribute levels in the DCE ...... 51 3.4 Comparison of samples 1 and 2 ...... 55 3.5 Swait-Louviere test results for certain vs. uncertain and consistent vs. inconsistent respondents ...... 60 3.6 Swait-Louviere test results for equality of choice behavior between samples 1 and 2 ...... 60 3.7 Swait-Louviere test results for equality of choice behavior between samples that differ in the position of the repeated choice task ... 61 3.8 Logit regression results ...... 63

4.1 Attributes and attribute levels in the choice experiment ...... 78 4.2 Sociodemographic characteristics of the study sample and target population ...... 84 4.3 Visual ANA statistics ...... 86 4.4 Full attribute-attendance and stated ANA MXL models ...... 88 4.5 Visual ANA MXL models based on different thresholds of lookup frequency and duration ...... 90 4.6 Inferred ECLC ANA model ...... 92 4.7 Shares of respondents displaying stated, visual, and inferred ANA 93 4.8 Baseline ECLC model ...... 100 xviii

5.1 Attributes and attribute levels in the DCE ...... 109 5.2 Explanatory variables included in the choice models ...... 114 5.3 Sociodemographic characteristics of the study samples and target population ...... 117 5.4 Perceived controllability of the comparative endpoint and middle point risks ...... 118 5.5 Estimated mixed logit models ...... 119 5.6 MWTP estimates for the risk attributes ...... 124 xix

List of Abbreviations

AIC Akaike Information Criterion ANA Attribute Non-Attendance ASC Alternative-Speciﬁc Constant BIC Bayesian Information Criterion CV Contingent Valuation DCE Discrete Choice Experiment ECLC Equality-Constrained Latent Class FAA Full Attribute Attendance fMRI functional Magnetic Resonance Imaging GWyr Gigawatt-year HTCM Hypothetical Travel Cost Method IEA International Energy Agency LC Latent Class LR Likelihood Ratio MNL Multinomial Logit MWTP Marginal Willingness-To-Pay MXL Mixed Logit OECD Organization for Economic Co-operation and Development PPP Purchasing Power Parity RP Reference Point SFOE Swiss Federal Ofﬁce of Energy SP Stated Preference SQ Status Quo TCM Travel Cost Method WTP Willingness-To-Pay

Chapter 1

Introduction

1.1 Background

Two trends are currently reshaping the world’s energy landscape: a shift from carbon-based to non-carbon-based sources of energy, and a diminishing role of nuclear power in global energy production. The ﬁrst trend is observable in the fact that renewable energy sources (excluding hydropower) have experienced double-digit growth rates since 2005 and doubled their share in global power generation within the last 5 years to 6.7% in 2015 (PBL, 2016). Their share in global primary energy consumption has also doubled since 2010, reaching 2.8% in 2015. This is reﬂected in a decreasing global growth rate of CO2 emissions over the last 15 years: While the 5-year average values for global annual emissions increased by 3.1% between 2001 and 2005, they grew by 2.5% and 1.4% in the years 2006 to 2010 and 2011 to 2015, respectively. In 2015 and 2016, global CO2 emissions growth rates came to a standstill and emissions remained largely unchanged (IEA, 2017; PBL, 2016). This may signal a decoupling of CO2 emissions and economic activity, as the world economy grew steadily in both years, with GDP growth rates of 3.2% and 3.1% for 2015 and 2016, respectively (IEA, 2017;

IMF, 2017). Nevertheless, the regional differences in the CO2 emission trends are considerable. In 2015, the United States and China reduced their CO2 emissions by 2.6% and 0.7%, respectively, while the European Union and India increased their emissions by 1.3% and 5.1%, respectively. The second global trend in energy supply is a shift away from nuclear power. The share of nuclear power in global electricity production peaked in the mid- 1990s at around 20% and has been declining since, reaching 11% in 2015 (IAEA, 2 Chapter 1. Introduction

2015; PBL, 2016). Nevertheless, energy policies with regard to nuclear power diverge in different parts of the world: Germany, Switzerland, and Belgium have decided, with various degrees of legally binding decisions already taken, to phase-out nuclear power. At the same time, a number of countries are expanding nuclear electricity production substantially, most notably China, Russia, and South Korea. In response to these trends, and coupled with a weakened public acceptance of nuclear power after the nuclear accident in 2011 in Fukushima, Japan, Switzer- land revised its energy policy in a strategic long-term report – Energy Strategy 2050 (Prognos, 2012; SFOE, 2013). The key components of the Swiss Energy Strategy 2050 include: the expansion of renewable sources of energy including hydropower; an increase in energy efficiency; and the phasing-out of nuclear power. This last aspect has important implications for the country’s electricity supply, which has primarily been based on hydropower and nuclear power over the last decades. Hydropower and nuclear power produced, respectively, 59% and 33% of total electricity in 2016 (SFOE, 2017). Nuclear power is intended to be replaced with, among others, renewable energy sources. With respect to hydropower, the Energy Strategy 2050 foresees an increase in electricity production of ca. 9% compared with the current hydropower production (SFOE, 2012). However, the hydropower industry in Switzerland is in a difficult situation for a number of reasons. Most of the hydropower plants currently operate with losses, as a result of considerably lower prices and a reduced variation between peak and off-peak prices on electricity markets in the recent years. There are three main reasons for this: The increasing share of renewable electricity generation; low global coal prices; and the low CO2 emission permit prices of the EU trading system (Barry et al., 2015). Although the environmental benefits of hydropower, such as low greenhouse gas emissions, are evident, the hydropower industry in Switzerland has a history of public dispute due to its environmental externalities. The negative effects of hydropower are mainly caused by a loss of connectivity between aquatic systems and altered flow regimes (hydropeaking). A loss of connectivity between water bodies affects the migration of animal species. Changes in flow regimes have an impact on wildlife, may endanger floodplains, and cause erosion. Another factor that hampers the industry is the 1.2. Main objective, hypotheses, and research questions 3 low public acceptance of the hydropower technology in Switzerland and elsewhere (Sternberg, 2008, 2010; Tabi and Wüstenhagen, 2017). This is likely to be associated with the above-mentioned negative environmental externalities of hydropower. Switzerland has the highest proportion (88%) of already developed hydropower potential worldwide (IEA, 2010). In addition, considering the economic, environmental, and social obstacles that the hydropower industry in Switzer- land currently faces, the envisaged expansion of hydropower production by 9% from its current level is challenging. The Swiss Federal Office of Energy hence concludes that such an expansion is only possible under a considerable improvement in economic conditions and public acceptance (SFOE, 2012).

1.2 Main objective, hypotheses, and research questions

This PhD thesis focuses on the public acceptance of hydropower expansion. It does so by investigating the values that the general public attaches to the various positive and negative externalities related to hydropower. These values are typically elicited using stated preference research methods, that is, research in which the public at large is asked to answer questions about the externalities involved, using different forms of survey methods. In the past two decades, discrete choice experiments (DCE) have become the most important stated preference (SP) elicitation approach (Johnston et al., 2017). The general objective of this PhD thesis is to quantify and explain variation in the non-market values at- tached to hydropower externalities. More specifically, it aims to test axioms and assumptions of microeconomic choice theory in the context of DCEs. A quantitative meta-analysis summarizes and synthesizes the results of the existing SP literature on hydropower externalities, and generates new insights that are used as inputs in the design of a DCE in this thesis. The choice survey that is subsequently designed and implemented consists of four versions which differ with respect to their methodological characteristics. Four independent samples of respondents answered the different versions of the DCE. This setup allows this dissertation to investigate the choice behavior of survey participants in a DCE and, in particular, it enables it to gain new insights into choice certainty, choice 4 Chapter 1. Introduction consistency, and choice monotonicity, as well as the continuity of preferences and their dependence on reference points. All of these concepts are linked to the axioms and assumptions of choice theory. This dissertation is embedded in the context of Swiss energy policy, as the DCE elicits public preferences and willingness-to-pay (WTP) values for an expansion of electricity production by hydropower in Switzerland. In the first part of this dissertation, a meta-analysis is conducted by estimating meta-regression models using the estimated non-market values for hydropower externalities. These values are elicited from existing SP studies implemented in different parts of the world. This is the first meta-analysis in the non-market valuation literature that explicitly focuses on hydropower and its external effects. Next, a DCE is conducted as part of an online survey. The policy context described in the previous section serves as a case study for the implementation of the DCE. Research on the external effects of renewable energy typically investigates either direct externalities, e.g. the effects on wildlife, vegetation, and landscape, or indirect externalities, e.g. greenhouse gas emissions. Greenhouse gas emissions constitute an indirect externality, because they are not directly caused by renewable energy sources but represent avoided external effects of conventional sources of electricity. In the Swiss energy context, one of the major indirect externalities of renewable energy sources is the avoidance of nuclear risk. Until now, the effect of avoiding nuclear risk on public risk perception and preferences for renewable energy has not been studied in the valuation literature. Hence, this PhD thesis contributes to the existing literature by its central hypothesis, which postulates that public preferences for expanding hydropower production are linked to public preferences for avoiding nuclear risk. To this end, a choice attribute related to nuclear risk is included in the DCE. Hartmann et al. (2013) is the only study that explicitly investigates the relationship between public attitudes and perceptions of nuclear power and preferences for the adoption of green electricity. However, in contrast to Hartmann et al. (2013), this thesis focuses on this relationship based on stated preference research. The main objective of the DCE is to investigate a number of axioms and common assumptions made in consumer choice theory (e.g. Jehle and Reny, 2001; Mas-Colell and Whinston, 1995). The underlying theoretical framework of DCEs 1.2. Main objective, hypotheses, and research questions 5 is provided by Lancaster’s theory of consumer demand (Lancaster, 1966), and by random utility theory (McFadden, 1974; Thurstone, 1927). Lancaster’s theory of demand states that utility is not derived from goods directly, but from their individual characteristics. Random utility theory has its roots in the rapidly increasing availability of survey data in the 1960s, which created the need for linking observed behavior to existing microeconomic consumer theory (McFadden, 2001). Random utility theory assumes that utility functions exist, and that respondents choose in accordance with their utility functions. Therefore, the standard economic axioms and assumptions of consumer theory as outlined, for example, in Jehle and Reny (2001), are maintained in the random utility framework. This PhD thesis tests some of these axioms and assumptions. Additionally, it tests an important assumption about choice behavior in Kahneman and Tversky’s (1979) prospect theory. First, the standard economic axiom of monotonicity and the conventional assumptions of stable and known preferences are investigated. This is accom- plished by testing the null hypothesis that preferences are known, consistent, and monotonic, and by examining the determinants of choice certainty, choice consistency, and choice monotonicity and their impact on choice behavior. The contribution of this PhD thesis to the existing literature is that this thesis investigates choice certainty, consistency, and monotonicity simultaneously, using the same choice responses. This allows for the identification of both common and idiosyncratic drivers of these constructs. Second, the axiom of continuous preference relations is tested by analyzing attribute non-attendance (ANA). Continuous preference relations imply continuous indifference curves, and hence assume that choice participants adopt compensatory decision-making rules when making choices in a DCE (e.g. Lagarde, 2010). The existence of ANA violates the continuity axiom. This thesis assesses how a novel methodology to capture visual ANA can contribute to a better understanding of ANA behavior. Specifically, the common approach for analyzing ANA in the existing literature through stated or inferred ANA information is extended with a novel, visual approach for capturing ANA behavior based on mouse-tracking. Finally, this dissertation studies the dependence of preferences on reference points, which is an important assumption of choice behavior in prospect theory. 6 Chapter 1. Introduction

In contrast to large parts of the valuation literature, the original text of Kahne- man and Tversky (1979) states a variety of possible reference points that do not have to coincide with an individual’s status quo. Drawing on prospect theory, this PhD thesis examines the possibility that comparative risks displayed on a risk ladder may serve as reference points and have an impact on an individual’s choice in a DCE. The majority of the DCE literature on reference points focuses on reference points that are linked to the characteristics of the choice tasks (e.g. their baseline levels). In contrast, the last chapter of this thesis contributes to the DCE literature by exploring the role of reference points which are induced independently and prior to the actual choice tasks. The main research questions addressed in the thesis can be summarized as follows:

1. What are the main determinants of the non-market values for hydropower externalities?

2. What are the common and idiosyncratic determinants of choice certainty, choice consistency, and choice monotonicity in DCEs, and what is the role of choice complexity?

3. How does visual ANA data obtained from mouse-tracking perform in explaining ANA behavior compared with stated and inferred ANA?

4. Do comparative risks shown on risk ladders serve as reference points and inﬂuence preferences for a change in risk?

1.3 Data collection and econometric analysis

Two data collection processes took place for the purpose of this PhD thesis. First, a database was created based on secondary data derived from existing SP studies and publications for the meta-analysis, and second, a survey including a DCE was conducted to collect primary research data. A database of existing research that values the external effects of hydropower was constructed in order to identify the main determinants of the economic values for hydropower externalities (research question 1). The created database consists of 29 international studies, which together generate 81 observations. Three different meta-regression 1.3. Data collection and econometric analysis 7 models are applied. A baseline model is estimated using weighted least squares regression analysis. The observations are weighted by the sample size of the surveys in the original studies. This procedure controls for differences in the variances of the values in the database by assuming that variances are smaller for observations that are obtained from surveys with larger sample sizes. Two other models control for systematic differences in mean welfare estimates between the studies, and for differences in the influence of the regressors on the dependent variable. The DCE that follows the global meta-analysis elicits public preferences for an expansion of hydropower in Switzerland specifically, and aims to generate data on public preferences for hydropower expansion. This serves to answer the methodological research questions 2 to 4. The DCE was implemented among 1,000 households that constituted a representative sample of the German- and French-speaking Swiss population (roughly 95% of the total population of Switzer- land, the remaining 5% live in an Italian-speaking region). Survey pretesting included 20 face-to-face interviews and two rounds of online pretests with 220 and 350 respondents. For the final DCE, the respondents were split into four different, independently recruited representative samples, each comprising ca. 250 households. Each sample received a slightly different questionnaire version in order to be able to answer the different research questions. Compared with the baseline version, the three other versions differed with respect to: the presence of questions on choice certainty; the monitoring of the respondents’ information acquisition process by mouse-tracking; and the reference points included in a risk ladder representing the changes in hydropower and nuclear power risk under valuation. Different econometric techniques are applied. For the purpose of answering question 2, binary logit models and random-effects ordered logit models are estimated. Binary logit models are used to regress choice consistency and choice monotonicity on possible explanatory variables, and random-effects ordered logit models are employed to identify drivers underlying stated choice certainty. To answer research questions 3 and 4, mixed logit (MXL) models are estimated. In contrast to the fixed effects multinomial logit model, MXL models allow for random taste heterogeneity across individuals and correlation between unobserved factors over alternatives and choice tasks. Question 3 is addressed 8 Chapter 1. Introduction by running MXL models with attribute parameters for respondents who state non-attendance to an attribute restricted to zero. The same procedure is applied for analyzing visual ANA information. Inferred ANA is assessed using equality- constrained latent class (ECLC) models. Each class in this model describes a specific pre-defined pattern of ANA behavior. MXL models are also used in order to identify the effects of different risk ladders on choice behavior (research question 4). The Swait and Louviere (1993) test procedure is applied in the course of answering the research questions 2 and 4. For research question 2, the tests assess whether there are statistically significant differences between the choice behavior of respondents who are (un)certain about and (in)consistent in their choices. Furthermore, the effect of including follow-up questions on choice certainty and including the same choice task in a different position in the choice-task sequence is examined using the same procedure. In answering research question 4, the Swait-Louviere test is applied to compare two split-samples of respondents who were shown risk ladders that differ with respect to the ranges of probabilities of comparative risks.

1.4 PhD thesis outline

Chapter 2 aims to answer research question 1. This study was first presented at the 22nd Annual Conference of the European Association of Environmental and Resource Economists in Zurich in June 2016, and has been published as Mattmann, Logar, and Brouwer (2016a) in Energy Economics. It presents a meta- analysis of existing SP research on the economic value of the positive and negative external effects of hydropower. For this purpose, a database with the economic values of the non-market impacts of hydropower electricity generation is constructed. The main aim of the meta-analysis is to quantify and explain the economic values for positive and negative hydropower externalities. Different meta-regression model specifications are estimated to test the robustness of the determinants of these non-market values. The impact of key methodological features of the valuation studies on the results is also investigated. Chapter 3 attempts to answer research question 2. It focuses on the consumer theory axiom of monotonicity and the assumptions that consumer preferences 1.4. PhD thesis outline 9 are known and stable. More specifically, Chapter 3 tests whether choices are based on known, stable, and monotonic preferences, and investigates the common and idiosyncratic determinants of choice certainty, consistency, and monotonicity based on the results of the DCE. For this purpose, choice certainty, consistency, and monotonicity are regressed on possible drivers. In doing so, two different measures of choice task complexity are compared: The entropy of a choice task, and the utility difference between the alternative that is chosen and the second-best alternative. Moreover, this chapter tests the equality of choice behavior of respondents who differ with respect to choice certainty and consistency. It also investigates the effect of including choice certainty follow-up questions after each choice task, and compares the choice behavior of respondents who were shown a repeated choice task in a different position in the choice task sequence. Chapter 4 answers research question 3 and investigates the standard economic axiom of continuity. This chapter was first presented at the 23rd Annual Confer- ence of the European Association of Environmental and Resource Economists in Athens in June 2017. It presents the first application of mouse-tracking to analyze ANA in DCEs. Mouse-tracking is applied to record the frequency and duration of uncovering attribute information in the choice process. Mouse-tracking functions similarly to eye-tracking, but can be applied online and allows for a larger sample size. The information obtained from mouse-tracking is used to generate a visual definition of ANA, while stated ANA information is collected by means of a follow-up question after the DCE. The performance of choice models based on stated, inferred, and visual ANA information is compared. Chapter 5 focuses on research question 4 and assesses a key assumption of prospect theory: the dependence of preferences on reference points. This chapter argues for the existence of multiple reference points. The DCE values the changes in the risk of dying caused by a hydropower and a nuclear power accident. Risk ladders are used to communicate the risk information to respondents. Two different risk ladders are presented to two independent samples of respondents. The risk ladders differ with respect to the range of risk probabilities that serve as a benchmark for the risks being valued. One sample is shown a risk ladder with a high reference point, that is, a risk ladder with a wide range of comparative risk probabilities that include high risk events, whereas the other 10 Chapter 1. Introduction sample is shown a risk ladder with a low reference point, i.e. a risk ladder with a narrow range of comparative risk probabilities that encompass lower risks. On both risk ladders, the change in risk that is being valued is identical for both samples. Chapter 5 hypothesizes that, in addition to the status quo probability of the valued risks, comparative risks presented on the risk ladders represent reference points that influence the valuation of risk changes. Chapter 6 discusses the results that are presented in the Chapters 2, 3, 4, and 5 and concludes. This chapter also identifies the need for further future research, and summarizes policy-relevant insights. 11

Chapter 2

Hydropower Externalities: A Meta-Analysis1

2.1 Introduction

As a result of both, increasing efforts to decarbonize economies and substantially diminished social and political acceptance of nuclear energy production following the 2011 accident in Fukushima, Japan, renewable energy sources are set to play a more prominent role worldwide. This is reﬂected in various national energy policies. Germany and Switzerland, for example, decided to phase out nuclear energy production and replace its share in national electricity production primarily with renewable energy sources (SFOE, 2013). Renewable energy sources avoid many negative externalities of conventional energy production based on fossil or nuclear fuels, which typically involve long-term consequences such as the impacts of greenhouse gas emission on climate change or the accumulation of radioactive waste. However, renewable sources of energy often operate with lower energy densities than non-renewable energy carriers, which results in spatially larger production facilities (Wüstenhagen, Wolsink, and Bürer, 2007). As a consequence, other types of externalities such as threats to biodiversity or aesthetic impacts occur.

1This chapter is published as: Mattmann, Matteo, Ivana Logar, and Roy Brouwer (2016). "Hy- dropower Externalities: A Meta-Analysis". Energy Economics 57, pp.66-77. It was also presented at the 22nd Annual Conference of the European Association of Environmental and Resource Economists in Zurich in June 2016. 12 Chapter 2. Hydropower Externalities: A Meta-Analysis

Much of the existing research related to the economic valuation of renewable energy focuses on the newer technologies of wind, solar, biomass and bio- fuel. Recent examples include studies which value externalities from: wind power generation (Alvarez-Farizo and Hanley, 2002; Ek, 2006; Ek and Pers- son, 2014; Ladenburg and Dubgaard, 2007); biomass (Susaeta et al., 2011); or a mixture of various renewable energy sources (Bergmann, Colombo, and Han- ley, 2008; Bergmann, Hanley, and Wright, 2006; Komarek, Lupi, and Kaplowitz, 2011; Kosenius and Ollikainen, 2013; Ku and Yoo, 2010; Longo, Markandya, and Petrucci, 2008). In contrast, the amount of research that has been conducted on the effects and economic values of more established technologies such as hydropower is rather limited. Since the role of hydropower as a source of renewable energy is expected to expand further worldwide (e.g. Jacobson and Deluc- chi, 2009), an understanding of individuals’ preferences for its effects on the environment, recreational activities, and aesthetic values is of crucial importance to inform an effective and efficient energy transition. Hydropower is a renewable source of energy with a long history (Paish, 2002). The product of hydropower generation is electricity, a standard market good that can be sold directly to electricity consumers, and it is therefore usually not considered in valuation studies. The same holds for the employment effects of hydropower operations. However, hydropower electricity production typically generates a number of positive and negative side effects that affect different groups of stakeholders, for which they are, in most cases, not (directly) compensated. These effects of hydropower depend not only on the size of operation and the geographical location, but also on the type of hydropower facility. That is, run-of-the-river facilities, usually operating with constant water flows and generating electric base load, have different effects than storage plants that depend on dams to store water, which is released at times of peak demand. The effects of storage plants with natural water feeding can differ again from the effects of pumped-storage plants that pump water from a lower to a higher reservoir. In general, most of the external effects of hydropower are caused by hydropeaking and disconnected water bodies. Reduced connectivity refers to the disconnec- tion of water bodies caused by hydropower dams and run-of-the-river facilities. Changes in flow (hydropeaking) occur only in the case of storage hydropower plants. Hydropeaking causes non-natural flow patterns, i.e. high variability in 2.1. Introduction 13 discharge, water levels, and flow velocity of water bodies. The various effects caused by different types of hydropower plants will be briefly summarized below. Recreation is an important service provided by aquatic ecosystems (Boyd and Banzhaf, 2007), which may be impaired by hydropower. Examples of such services affected by hydropower operations include various types of recreational activities such as kayaking or rafting (Aravena, Hutchinson, and Longo, 2012; Hynes and Hanley, 2006), fishing (Filippini, Buchli, and Banfi, 2003; Gogniat, 2011; Håkansson, 2009; Loomis, Sorg, and Donnelly, 1986; Navrud, 2004; Rob- bins and Lewis, 2009) or visiting waterfalls (Ehrlich and Reimann, 2010). Most studies observe that these recreational activities are negatively influenced by hydropower due to hydropeaking and the disconnectivity of water bodies, both of which impede water sports and endanger fish populations, thereby reducing the value of angling. It is, however, conceivable that hydropower may also generate positive effects on recreational opportunities: for example, by creating artificial lakes suitable for water sports. Getzner (2015) empirically compares the recreational value of free-flowing sections of a river with dammed stretches and finds higher recreational benefits on free-flowing sections than on dammed stretches of rivers for a variety of recreational activities. The environmental effects of hydropower are manifold. A positive environmental externality of hydropower electricity production is lower greenhouse gas emission compared with most other sources of electricity production (see Weisser (2007) for a literature overview of greenhouse gas emissions by different electricity production technologies). The reduction in the emission of greenhouse gases depends, however, on reservoir size and type, the extent of flooded vegetation, soil type, water depth, and climate conditions. Especially methane emission can form a significant source of greenhouse gas release in the case of the hydropower reservoirs of storage plants in tropical regions (e.g. Barros et al., 2011; Delsontro et al., 2010). Pumped-storage plants without natural water feed are used for load balancing only, and do not directly reduce greenhouse gas emissions since they consume more electricity than they generate. Negative environmental externalities of hydropower also stem from either reduced connectivity of aquatic systems or altered flow regimes. Reduced connectivity especially affects the migration of fish and other animal species. Changes 14 Chapter 2. Hydropower Externalities: A Meta-Analysis in flow patterns (hydropeaking) change sedimentation levels, and can lead to rapid changes in water temperature. Both of these effects have an impact on invertebrates, which are usually very sensitive to altered temperature and sediments (e.g. Bruno et al., 2009). In addition, non-natural hydropower flow patterns may endanger floodplains, threaten fish and bird species, and cause erosion. Hydropower projects, especially the construction of dams, artificial lakes and reservoirs, may also affect artifacts of important cultural, historical and geolog- ical value that are flooded during the construction phase of hydropower storage plants (Han, Kwak, and Yoo, 2008; Lienhoop and MacMillan, 2007; Navrud, 2004). Direct, potentially negative, aesthetic impacts of hydropower often stem from hydropower-related facilities such as dams, access tracks, pipelines, build- ings, and the lack of vegetation due to these installations (Hanley and Nevin, 1999). Run-of-the-river plants cause aesthetic degradation as well. It has been shown that free-flowing rivers have higher aesthetic value compared with rivers affected by hydropower facilities (Born et al., 1998). Furthermore, pylons connecting remote hydropower plants might adversely affect views and scenery (Aravena, Hutchinson, and Longo, 2012). The main objective of this paper is to synthesize the empirical evidence on the economic valuation of hydropower externalities in a meta-analysis. In contrast to a recent meta-analysis on the willingness-to-pay for green electricity (Sundt and Rehdanz, 2015), we focus explicitly on hydropower and its externalities. This is, to our knowledge, the first study to conduct such an analysis. A main research question addressed in this paper is whether the positive hydropower externalities outweigh the negative ones. The purpose of the meta-analysis is not only to review and evaluate the existing literature, but also to explain study-to-study variation by focusing on differences between valuations for various positive and negative types of hydropower externalities, as well as on key methodological characteristics such as sensitivity to scope. In order to do this, the external effects of hydropower production are first identified and classified. Next, the drivers of welfare estimates for the non-market effects of hydroelectric production technology are examined in a meta-regression model. The remainder of this paper is structured as follows. Section 2.2 describes the search procedure and selection of studies included in the meta-analysis. Section 2.2. Study selection and characteristics 15

2.3 explains the main econometric issues in meta-modeling and the estimated models. Section 2.4 considers the factors that inﬂuence the economic values of hydropower externalities. The results of the estimated meta-regression models are presented in Section 2.5 followed by conclusions in Section 2.6.

2.2 Study selection and characteristics

The non-market valuation of the externalities of hydropower production constituted the main criterion for a study to be included in the meta-analysis. More specifically, all studies that generated primary valuation data of the non-market impacts of electricity production by hydropower were considered for inclusion. We included all studies in which hydropower production was identified as a source of the externalities. This involves studies that valued the externalities of hydropower exclusively (roughly 80% of all observations), as well as studies which value the external effects of renewable energy in general, but explicitly mention hydropower to be one of these (20% of the observations included). For example, a study that values increased water flows due to modified hydropower operation schemes would be included in the analysis, whereas a study that estimates the value of increased water flows without explicitly specifying that these changes in water flows are caused by hydropower operation would not be included. Applying this selection criterion ensured that individuals took their preferences for hydropower into account when valuing the external effects. The search procedure was conducted in 2014. Online databases that were browsed included Google Scholar, Scopus, Econlit and RePEc. ProQuest was used to search specifically for relevant PhD theses. The search included both published and unpublished papers, working papers, conference papers, PhD theses, Master’s theses, government and non-government reports. Keywords that were used in the search process included, among others, the following terms and combinations thereof: hydropower, hydroelectric, stated preferences, revealed preferences, contingent valuation, conjoint analysis, choice experiment, travel cost, hedonic pricing, externalities, dams, and recreational benefits. Table 2.1 provides a list of the studies included in the meta-analysis collected by the search and selection procedures described above. Most of the studies obtained are articles published in international peer-reviewed journals, but there 16 Chapter 2. Hydropower Externalities: A Meta-Analysis are also two reports, two working papers, one conference paper, a PhD thesis, and two Master’s theses. Three reports could not be obtained despite an extensive search procedure. Other studies that were excluded to avoid double counting analyzed data that had already been used in one or more other relevant publication. Five papers valued the externalities of renewable energy in general without explicitly mentioning hydropower, and thus the economic values of the effects could not be ascribed to hydropower. Furthermore, two publications reported only aggregated economic values for the relevant population that could not be transformed to individual welfare estimates. The earliest study was carried out in 1983, while the other studies were conducted over a period of 18 years between 1993 and 2011. The majority of the studies were carried out in Europe (70%), followed by South America (13%), the United States (9%), and Asia (9%). With respect to the valuation methods, most studies applied stated preference methods, such as contingent valuation (CV) or discrete choice experiments (DCE); two studies used revealed preference methods (travel cost method (TCM)); and three combined revealed and stated preference approaches, using the hypothetical TCM (HTCM). Out of a total of 29 studies, 81 observations could be used in the subsequent meta-analysis. 15 studies contributed only one observation. Studies provided more than one observation when using different samples of respondents (for example, distinguishing between users and non-users of a resource) or because they valued various combinations of hydropower externalities. A few studies also varied the methodological aspects in split samples. The number of respondents underlying each observation varies considerably (between 45 and 1933), with an average of 361 respondents per observation. Eight observations (9.9%) included respondents who were directly affected by hydropower externalities. These are, for example, anglers, who were asked to value the number of fish in a river affected by hydropower. Peer-reviewed papers included in the analysis received, on average, 39 citations measured by the Google Scholar citation index, with one study having a maximum of 136 citations (up to December 2014). Finally, the share of hydropower in total national electricity production (in the year of the survey) was included as a measure for the energy mix in a country (IEA, 2014a,b). Na- tional shares of hydropower vary widely, with an average of 38% of electricity produced by hydropower in the countries where the surveys were conducted. 2.2. Study selection and characteristics 17

TABLE 2.1: Studies collected in the selection and search procedure (ordered by study year)

# Study Authors Type of publication Country Nat. Valuation Nc year (years of hydro. Methodb publication) sharea

1 1983 Loomis, Sorg, Journal article (Journal USA 13.7% CV 1 and Donnelly of Environmental (1986) Management) 2 1993 Kosz (1996) Journal article AUT 71.5% CV 1 (Ecological Economics) 3 1993 Navrud Report & Journal NOR 99.6% CV 2 (1995, 2001) article (Hydropower and Dams) 4 1994 Biro (1998) Journal article (Ambio) TUR 39.1% CV 1 5 1996 Loomis (1996) Journal article (Water USA 9.6% CV 3 Resources Research) 6 1997 Hansesveen Master’s Thesis NOR 99.3% CV 3 and Helgas (1997) 7 1998 Bergland Report NOR 99.4% CV 3 (1998) 8 1998 Filippini, Journal article CHE 53.7% HTCM 1 Buchli, and (Applied Economics) Banﬁ (2003) 9 1998 Hanley and Journal article (Energy GBR 1.4% CV 1 Nevin (1999) Policy) 10 1998 Loomis (2002) Journal article (Water USA 7.8% HTCM 1 Resources Research) 11 2002 Han, Kwak, Journal article KOR 1.0% DCE 1 and Yoo (Environmental Impact (2008) Assessment Review) 12 2002 Sundqvist Doctoral Thesis SWE 45.2% DCE 1 (2002) 13 2003 Bothe (2003) Working Paper ISL 83.4% CV 1 14 2003 Hynes and Journal article (Land IRL 2.4% TCM 1 Hanley (2006) Use Policy) 15 2003 Bergmann, Journal article GBR 0.8% DCE 6 Colombo, (Ecological Economics) and Hanley (2008) 16 2004 Håkansson Journal article (Journal SWE 39.6% CV 8 (2009) of Environmental Planning and Management) 17 2004 Navrud Report NOR 98.8% CV 1 (2004) 18 Chapter 2. Hydropower Externalities: A Meta-Analysis

# Study Authors Type of publication Country Nat. Valuation Nc year (years of hydro. Methodb publication) sharea

18 2005 Longo, Journal article GBR 40.5% DCE 4 Markandya, (Ecological Economics) and Petrucci (2008) 19 2006 Kataria (2009) Journal article (Energy SWE 43.1% DCE 7 Economics) 20 2006 Robbins and Journal article (Journal USA 6.8% TCM 2 Lewis (2009) of the American Water Resources Association) 21 2006 Ku and Yoo Journal article KOR 0.9% DCE 3 (2010) (Renewable and Sustainable Energy Reviews) 22 2008 Aravena, Journal article Energy CHL 40.5% CV 1 Hutchinson, Economics) and Longo (2012) 23 2008 Ponce et al. Journal article (Water CHL 40.5% CV 10 (2011) Resources Management) 24 2008 Kosenius and Journal article (Energy FIN 22.1% DCE 1 Ollikainen Policy) (2013) 25 2009 Ehrlich and Journal article EST 0.4% CV 1 Reimann International Journal (2010) of Geology) 26 2010 Guo et al. Journal article (Energy CHN 17.2% CV 2 (2014) Policy) 27 2011 Gogniat Master’s Thesis CHE 51.5% HTCM 1 (2011) 28 2011 Klinglmair Conference Paper AUT 55.0% DCE 3 and Bliem (2013) 29 2011 Klinglmair, Working Paper AUT 55.0% DCE 10 Bliem, and Brouwer (2012) Notes: a IEA (2014a,b). b CV: Contingent Valuation; CE: Choice Experiment; HTCM: Hypothetical Travel Cost Method; TCM: Travel Cost Method. c Number of observations included in the meta-analysis. 2.3. Meta-model 19

2.3 Meta-model

2.3.1 Treatment of heterogeneity, heteroskedasticity, and non- independence

Meta-regression models can be classified according to the way they address and treat data heterogeneity, heteroskedasticity of effect-size variances, and non- in- dependence of observations from the same studies (Nelson and Kennedy, 2008). This section explains these three issues and how they are tackled in our study. Data used in a meta-analysis come from a variety of papers, authors, and countries. Furthermore, there are often individual-specific differences between survey participants, and the effect-size that forms the dependent variable in a meta-analysis might suffer from inconsistencies between studies (Smith and Pat- tanayak, 2002). In other words, studies may differ with respect to their design elements, but they may also differ regarding their examined effect-size (Ringquist, 2013). Apart from enhancing the comparability of effect-sizes by adjusting available data from primary studies and dropping observations that lack comparability, the standard treatment of data heterogeneity in economic studies is to control for differences in effect-size by including independent variables (Nelson and Kennedy, 2008; Smith and Pattanayak, 2002). In this study, control will be included for the differences between the types of hydropower externalities valued, the sample characteristics, and the methodological features of different studies. The primary studies used in meta-analysis usually do not have the same (estimated) variances owing to differences in study-specific characteristics (Nelson and Kennedy, 2008). The standard assumption of the ordinary least squares (OLS) estimator of homogeneity is thus in general violated (Ringquist, 2013). In order to mitigate heteroskedasticity of effect-size variances and to control for differences in the quality of study results, the observations are ideally weighted by the inverse of their variances, resulting in weighted least squares regression (e.g. Lipsey and Wilson, 2001). By applying weights in this manner, more ac- curate studies with lower variances receive higher weights in the meta-analysis. Since, in this study, we only have information available about estimated variances of a fraction of the primary studies, we weight the individual observations by the square root of the study sample sizes, as is commonly done in the meta- regression literature (see Nelson and Kennedy (2008) for an overview on studies which apply this procedure). This ensures that studies with larger sample sizes (and therefore, as expected, also lower variances) receive more weight in 20 Chapter 2. Hydropower Externalities: A Meta-Analysis the analysis. As a consequence, the issue of heteroskedasticity is mitigated, and we ensure that the observations which we consider to be more reliable receive higher weights in the analysis. It is common procedure in meta-analysis to draw several effect-sizes from each study. Since observations drawn from the same study usually share some common characteristics, it must be assumed that there is within-study correlation between observations (Nelson and Kennedy, 2008). Various procedures exist to mitigate this issue, such as including only one observation per study, or including only mean values of various observations from the same study. How- ever, since the number of primary studies, and hence observations that are used in a meta-analysis, may be limited, it is in many cases unavoidable to use all the observations obtainable from each study. Furthermore, the use of several observations from the same study provides some estimation leverage because many elements of the research design of these observations remain the same (Ringquist, 2013). If various observations per study are used, it is necessary to control for within-study correlation by explicitly taking the hierarchical data structure into account. This can be done, for example, by using panel data models or calculating cluster-robust standard errors (Nelson and Kennedy, 2008). Both approaches are applied in this study.

2.3.2 The meta-regression models

We apply a variety of different approaches to address the issues described in Sec- tion 2.3.1, resulting in three different models. In Model 1 we use cluster-robust standard errors, where studies are set as clusters. This enables us to take the correlation between value estimates from the same studies into account. Cluster- robust standard errors assume independent observations across, but not within, clusters. Model 2 is a random-effects panel model with individual studies defined as cross-sectional units. Model 3 is an extension of the random-effects model that allows not only intercept coefficients, but also slope parameters to be random (Cameron and Trivedi, 2005). The baseline model (Model 1) is estimated by weighted least squares, and is specified as follows (e.g. Harbord and Higgins, 2008):

2 yi = xi0 + "i with "i [0, ], (2.1) ⇠ wi 2.3. Meta-model 21

where yi denotes the dependent variable, i.e. the welfare estimates for hydropower externalities; xi is a vector of regressors; and is a vector of associated coefficients. The observations are weighted by the square root of their respective sample size in this model. This was incorporated by using analytic weights which 2 assume an error term with mean equal to zero and weighted variance of /wi, 2 where is an unknown variance estimated in the regression, and wi denote the known weights. Variances are assumed to be smaller for observations that are based on a larger sample size. Cluster-robust standard errors are applied in order to control for within-study correlations of observations. The model above serves as a baseline case and is compared with more elaborate models (Models 2 and 3). Despite the advantage of the fixed-effects model that allows for correlation between unobservable study-specific effects and independent variables, such a specification is not feasible in our case because there is a substantial number of studies which provide only one observation. Model 2 therefore incorporates random-effects and is estimated using the maximum likelihood estimation procedure:

2 2 y = x0 + µ + " with µ [µ, ] and " [0, ]. (2.2) ij ij j ij i ⇠ µ ij ⇠ "

Model 2 incorporates two error terms: "ij denotes the standard error term, while

µj is a random variable that varies across j studies, but is assumed to be distributed independently of the regressors (Cameron and Trivedi, 2005). Both the random effects and the error term are assumed to be identically and independently distributed (iid). Model 2 is a more realistic speciﬁcation compared with Model 1 because it allows systematic differences in mean welfare estimates between studies to be captured. However, an even more elaborate model would make it possible to control for differences in the inﬂuence of regressors on the dependent variable between studies. Such differences can be modeled by taking into account not only random intercepts, but also random slope parameters. This results in a mixed-effects model (Model 3)2, which can be described as follows (Cameron and Trivedi, 2005):

2 y = x0 + z0 µ + " with " [0, ], (2.3) ij ij ij j ij ij ⇠ " 2Depending on the context of application, such models are also called hierarchical, multilevel, random coefﬁcients, or variance components models. 22 Chapter 2. Hydropower Externalities: A Meta-Analysis

where xij denotes, as before, the regressors; zij is a vector of observable characteristics (a subset of xij that includes the variables in the random part); µj is a random vector; and "ij is the standard error term. Mixed-effects models allow for the estimation of both fixed-effects and random-effects. Fixed-effects in this context describe the ordinary effects of regressors on the dependent variable. Their slope and intercepts describe the sample as a whole. These are the main effects of interest. Random-effects are the intercepts and slope parameters that vary across studies, and capture the heterogeneity between studies. Random- effects are usually not estimated directly, but their variances are calculated instead. The size and standard errors of these variances indicate whether there are significant variations between studies in the slope coefficients of the regressors that are assigned to the random part (Hamilton, 2012).

2.4 Selection and deﬁnition of variables

The main goal of the meta-analysis presented here is to explain variations in effect-size estimates between different studies, that is, variation in the dependent variable of interest, here welfare estimates for the positive and negative externalities associated with hydropower. The value function that serves as a conceptual basis for the different categories of factors which explain variations in effect-size estimates can be speciﬁed as follows:

W elfare estimatei = f(Qi; Ri; Si), (2.4) where the estimated economic value obtained from study i represents the effect- size of interest, i.e. the dependent variable whose variation we aim to explain. Q denotes the type of externality that is valued in study i. Of importance here is not only the externality itself, but also the size of change in the provision or quality level of the externality (i.e. the difference between Q1 in a new state and Q0 in the status quo). The various externalities (Q) valued were divided into the following five categories: (1) landscape and vegetation; (2) wildlife; (3) greenhouse gas emission; (4) recreation; and (5) aesthetics. Additionally, sample characteristics (R) and methodological features of the studies (S) are theoretically expected to play a significant role in explaining effect-size estimates. Sample characteristics (R) refer to the socio-economic characteristics of survey respondents, and methodological features (S) refer to the methods and procedures used to elicit and analyze the welfare estimates. 2.4. Selection and definition of variables 23

Table 2.2 provides a full list of the regressors included in the meta-regression model. The directions of the valued effects are also indicated, i.e. whether they describe improvements, mitigations, or deteriorations. Mitigations include policies such as restoring rivers or dismantling hydropower dams, all of which may mitigate the negative impacts of hydropower operation on landscape and vegetation, wildlife, recreation, and aesthetics. Mitigations thus describe the positive changes of negative hydropower externalities. In contrast, improvements refer to welfare measures for the positive changes of the positive externalities of hydropower. Since low greenhouse gas emissions are the only positive externality of hydropower valued in our data set, improvements refer exclusively to reducing greenhouse gas emissions. Negative changes in externalities, such as a negative change of aesthetic values, are described as deteriorations. The different directions of the valued effects are captured by separate externality-speciﬁc variables3. The dummy variables for deteriorations in landscape and vegetation and wildlife were merged into one variable due to perfect collinearity between the two (all observations that valued deteriorations in landscape and vegetation also valued deteriorations in wildlife).

TABLE 2.2: Explanatory variables included in the meta-analysis

Variables Description Coding of variables

Type of externality and size of change valued Landscape & Mitigation of negative impacts on Dummy: 1= Mitigation of Vegetation landscape & vegetation such as negative impacts on (mitigation) forests, flora species, or river-margin landscape & vegetation vegetation valued; 0=Otherwise Landscape & Deterioration of landscape & Dummy: 1=Deterioration of Vegetation and vegetation as well as deteriorations of landscape & vegetation and Wildlife wildlife wildlife valued; (deterioration) 0=Otherwise Wildlife Mitigation of negative impacts on Dummy: 1= Mitigation of (mitigation) fauna, especially populations of fish, negative impacts on wildlife birds and invertebrates (e.g. valued; 0=Otherwise improving fish passage) Greenhouse gas Reduction of greenhouse gas emission Dummy: 1=Reduction of emission greenhouse gas emission (improvement) valued; 0=Otherwise

3No control was included for differences in welfare measures (compensating or equivalent surplus measures) due to multi-collinearity, although the direction of the valued effects does not necessarily coincide with these welfare measures. A mitigation of an effect, for example, can be assessed by both a compensating and an equivalent surplus, depending on whether the mitigation is framed as an actual improvement or an avoided deterioration. 24 Chapter 2. Hydropower Externalities: A Meta-Analysis

Variables Description Coding of variables

Recreation Mitigation of negative impacts on Dummy: 1= Mitigation of (mitigation) recreational amenities affected by negative impacts on hydropower production, e.g. recreation valued; kayaking, river rafting, hunting or 0=Otherwise visiting a waterfall Aesthetics Mitigation of negative visual impacts, Dummy: 1= Mitigation of (mitigation) such as visibility of access tracks, negative visual impacts pipelines and pylons or general valued; 0=Otherwise aesthetic perception of water bodies that are used for hydropower Aesthetics Deterioration of visual impacts such Dummy: 1=Deterioration of (deterioration) as visibility of access tracks, pipelines visual impacts valued; and pylons or general aesthetic 0=Otherwise perception of water bodies that are used for hydropower Size of change Variable describing the size of an Dummy: 1=Small change impact of a valued externality valued; 0=Medium or large change valued Methodological variables Valuation method Describes the valuation method Dummy: 1=DCE; 0=CV, applied: discrete choice experiment TCM or HTCM (DCE), contingent valuation (CV), or travel cost methods (TCM or HTCM) Survey mode Describes the survey administration Dummy: 1=Face-to-face mode, e.g. mail, mail & phone, online, survey; 0=Other survey or face-to-face survey mode (mail, mail & phone, online) Payment vehicle Characterizes the payment vehicle Dummy: 1=Increase in used, e.g. tax increase, electricity taxes; 0=Other payment costs, water costs, entrance fees, etc. vehicles Payment duration Variable describing the duration of the Dummy: 1=Payment payment that is presented to duration is limited (one-off participants in the valuation or one year); 0=Unlimited procedure payment duration (inﬁnite) Sample characteristics North and South Continent of survey implementation Dummy: 1=North or South America America; 0=Elsewhere Asia Continent of survey implementation Dummy: 1=Asia; 0=Elsewhere Hydropower share Share of hydropower in national Continuous variable (%) electricity production Users Describes whether participants in the Dummy: 1=Users; valuation exercise are direct users of 0=Non-users the resource being valued (mainly anglers) 2.4. Selection and deﬁnition of variables 25

Variables Description Coding of variables

High income Median disposable household income Dummy: 1=Income above of all studies in 2013 USD (adjusted the median of all studies; for GDP purchasing power parities) 0=Income below the median Year of study Year of survey implementation Continuous variable (1983-2011)

With respect to the dependent variable, only mean welfare estimates of DCEs, that is, welfare estimates for scenarios entailing combinations of changes in externalities to assess the trade-offs involved, can be compared with the values obtained from CV and TCM studies. Marginal estimates of welfare obtained from DCEs were therefore excluded from the analysis. Mean welfare measures may represent slightly different concepts, depending on whether stated or revealed preference methods are used (Hicksian or Marshallian surplus measures). How- ever, for low income elasticities of demand for the externalities valued (and there is some evidence that income elasticities of demand for environmental goods are below unity: see, for example, Hökby and Söderqvist (2003)), Marshallian and Hicksian measures of surplus are similar, and it is therefore considered reasonable to use both measures in the same analysis. The effect-size estimates of the various studies also had to be made comparable. For this purpose, all estimates of welfare were expressed in 2013 USD by adjusting for annual consumer price inflation and the GDP purchasing power parities (PPPs) of the countries where the studies were conducted (OECD, 2014). The same procedure was applied to the income variable. An additional important adjustment was to express all the welfare measures on an annual basis to the degree that this was possible. Welfare estimates obtained from publications that defined the payment vehicle as a payment "per trip" were adjusted by the average annual number of trips where possible, and excluded from the analysis otherwise. One observation defined the payment vehicle as an increase in electricity costs per kWh. Since the survey sample of this study is representative of the national population, we transformed this welfare estimate to an annual electricity cost, using the average kWh consumption per household per year in the country where the study was conducted (IEA, 2014a,b). In most of the studies, the duration of the payment was specified as indefinite. However, 24 observations include one-off payments or payments of limited duration (one, five, or ten years). To control for payment duration, a dummy variable which distinguishes between short- and long-term payments was created. Short-term payments are defined as payments with a duration of up to one year. All durations longer than 26 Chapter 2. Hydropower Externalities: A Meta-Analysis one year are subsumed in the dummy for long-term payments. This approach is supported by experimentally observed discounting strategies, such as hyper- bolic discounting, that suggest high mental discounting rates in the short run and low behavioral weight of the future (e.g. Kirby and Herrnstein, 1995). The size of the change that is valued and the related notion of sensitivity to scope is a key conceptual issue accounted for in our meta-analysis. Sensitivity to scope describes the existence (or lack) of variation in economic values due to changes in the magnitude of an environmental good being valued (Carson, 1997). Sensitivity to scope was identified as one of the crucial criteria for valid and reliable stated preference research by the NOAA Panel (Arrow et al., 1993). Although there is an extensive literature on this issue, the results are somewhat inconclusive, and it is not always evident what an adequate response to scope would be. In general, most of the research concerning the existence and impact of scope effects has taken place in CV studies (see Desvousges, Mathews, and Train (2012) and Ojea and Loureiro (2011) for meta-analyses of the existing literature). This is also because DCEs, in contrast to CV, implicitly test for scope effects. The size of the change variable included in this study distinguishes between small, medium and large changes. Classification of the size of change was done based on the baseline and policy scenarios descriptions provided in the individual studies. This classification is available from the authors upon request. Special care was taken in the process of selecting variables to ensure that the conceptually most-relevant variables are included in the meta-analysis, and at the same time multicollinearity is avoided. For example, the dummy for the TCM cannot be included in the regression model as it is highly correlated with the dummy for direct users of a resource because TCM assesses only the values of users. The dummy for users, therefore, also captures a large part of the effect of using the TCM. As a consequence, the dummy for DCEs (1 if a DCE is applied, 0 for other valuation methods) can be interpreted as capturing the effect of using DCEs compared with using CV only. Similarly, the dummy variable for recreational amenities excludes fishing, because recreational fishing is highly correlated with the dummy for users, i.e. anglers in most cases. Dummy variables describing the payment unit (household versus individual), type of welfare measure (compensating or equivalent surplus), and cultural heritage values also caused multicollinearity issues (in addition to not turning out to be significant in any model), and were hence not included in the analysis. The same holds for the variable which tests for differences in values for the externalities of existing and planned hydropower facilities. 2.5. Results 27

2.5 Results

2.5.1 Descriptive statistics

Table 2.3 shows the cross-tabulation of the mean economic values across the main explanatory variables considered for the meta-regression models. The last row summarizes the welfare measure for the overall sample. Since a test of the equality of economic values between studies which value externalities of hydropower production exclusively and studies which include other renewable energies as well showed that these two welfare measures do not differ significantly, both categories of studies were included for the descriptive statistics and meta-regressions. Furthermore, no significant differences were found between the values associated with storage plants (27% of all observations), run-of-the- river plants (46%), and observations that do not distinguish between these different types of hydropower plants (27%). The estimates obtained from different hydropower types are therefore pooled in our analysis. It is not meaningful to disentangle the economic values estimated for different categories of externalities, since most of these values represent a combination of attributes. The results of the Kruskal-Wallis test indicate that the welfare measures differ significantly between continents, as well as between valuation methods. The mean values for different regions show that surveys conducted in North and South America result in a significantly higher PPP adjusted welfare value than in Europe or Asia. They also show that welfare estimates in Asia are generally the lowest. Note, however, that the number of observations in Asia is limited, and the standard error is high. The same applies to the relatively high values found for North and South America. Contrary to expectations, the TCM generates the highest values of the three valuation methods. Here also the results have to be interpreted with caution because of the low number of observations and the relatively high standard errors for TCM. The observed differences in welfare estimates between different survey administration modes, types of welfare measures and size of change categories are not statistically significant. The variable that captures the sensitivity to scope (i.e. the size of change) indicates that values increase when the size of change shifts from small to medium, but slightly decrease again for shifts from medium to large changes in externalities. This result might suggest insensitivity to scope or at least diminishing marginal utility of individuals when moving from small to medium and then to large impacts of hydropower. 28 Chapter 2. Hydropower Externalities: A Meta-Analysis

TABLE 2.3: Cross-tabulation of mean values of hydropower externalities across groups of explanatory variables

Mean Std. Min. Max. Na Kruskal- value Err. value value Wallis test (2013 statistic USD) Continents 2=6.60 p=0.04 North and South 275.1 401.9 87.7 1841.2 18 America Europe 146.9 164.3 3.9 1033.8 56 Asia 94.3 166.9 14.8 471.8 7 Valuation techniques 2=22.24 p=0.00 DCE 152.6 106.6 14.8 487.6 36 CV 97.6 94.1 3.9 471.8 39 TCM 732.7 603.7 337.1 1841.2 6 Survey administration 2=1.87 p=0.39 Face-to-face 131.5 105.8 10.9 471.8 35 Mail & mail/phone 215.3 361.0 3.9 1841.2 32 combined Online 167.7 99.7 15.1 370.6 14 Welfare measures 2=0.93 p=0.34 Compensating surplus 174.2 272.1 3.9 1841.2 61 Equivalent surplus 160.7 108.0 10.9 471.8 20 Size of change 2=0.51 p=0.78 Small 124.2 96.5 6.5 252.5 12 Medium 182.4 202.6 5.8 1033.8 31 Large 176.2 298.7 3.9 1841.2 38 Mean economic value 170.9 241.5 3.9 1841.2 81 Note: aNumber of observations. 2.5. Results 29

2.5.2 Meta-regression results

The dependent variable was adjusted using a Box-Cox power transformation in order to reduce its skewness (Box and Cox, 1964). The Box-Cox transformation estimates a parameter from the data that minimizes the skewness of the variable that is to be transformed (x):

x 1 if =0 B(x, )= 6 (2.5) ( ln x if =0.

By setting a specific value for , the Box-Cox transformation can incorporate many traditional transformations such as square, cubic or fourth root, as well as logarithmic transformations (Osborne, 2010). For example, =0 would indicate a natural logarithmic transformation to fit the data best. In our case the transformation of the dependent variable resulted in =0.17, implying that such a Box-Cox transformation is an even better fit for the data than a logarithmic transformation. Table 2.4 presents the outcomes of the three models described in Section 2.3.2. All models perform well with an R2 of 0.77 for the first model and a pseudo-R2 of 0.358 and 0.409 for Models 2 and 3, respectively. However, the pseudo-R2 lacks the explanatory power interpretation of the R2 for Model 1, and is therefore not directly comparable. Nevertheless, the Akaike information criterion (AIC) and the Bayesian information criterion (BIC) show that there is a slight improvement when moving from Model 2 to Model 3. The coefficients for the types of externalities confirm that the deterioration caused by hydropower production is valued, as expected, highly negative. This is evident from the significant negative coefficients for the deterioration of landscape, vegetation, and wildlife in all three models. The coefficients for aesthetic deterioration are negative in two out of three models but only reach significance in Model 3. Mitigating negative hydropower externalities does not seem to affect economic values substantially. The coefficients for mitigations of landscape and vegetation, as well as for wildlife, are not significant in two out of the three models. Furthermore, the coefficients for the deterioration of landscape, vegetation, and wildlife are much higher in absolute numbers than the estimates for mitigating these effects. Mitigation of aesthetic and recreational effects do not impact economic values significantly either. The coefficient for reducing greenhouse gas emissions through hydropower is not significant. However, when interacting the dummy for greenhouse gas 30 Chapter 2. Hydropower Externalities: A Meta-Analysis reduction with the share of hydropower in national electricity production, the coefficient of the interaction term is positive and highly significant. This means that reducing greenhouse gas emissions is valued positively and significantly more in countries with a higher share of hydropower in electricity production. A possible explanation for this result may be that awareness levels with respect to the positive effect of hydropower on greenhouse gas emission are higher in countries with a higher dependence on this renewable energy source. In order to assess the trade-offs between positive and negative externalities more quantitatively, an alternative version of Model 1 was estimated by applying a logarithmic transformation of the dependent variable. This produces qual- itatively similar results as shown in 2.4, but allows for a more straightforward interpretation of the different coefficients. According to this model specification, the deterioration of landscape, vegetation, and wildlife results, ceteris paribus, in a reduction of the estimated economic value by 136%. In contrast, the positive externality of avoiding greenhouse gas emissions in combination with the national hydropower share has a much weaker impact on the estimated non- market values. For each percentage point increase in the national hydropower share, avoiding greenhouse gas emissions results in a roughly 2.3% increase of the economic value. The relative change in the share of hydropower would have to be at least 60 percentage points in order to compensate for the valued negative externalities of hydropower production. Because the values of the medium and large specification of the scope variable are the same and not significantly different, a dummy variable is included for small changes only. The sign and significance of this variable in Models 2 and 3 provide evidence for economic values being sensitive to scope. The results obtained in this study therefore support the existing evidence on sensitivity to scope in the economic valuation literature (see, among others, Bateman and Brouwer, 2006; Carson, 1997; Carson and Mitchell, 1993; Ojea and Loureiro, 2011; Smith and Osborne, 1996). In contrast to most of the existing literature on scope sensitivity, and especially the comprehensive meta-analysis of Ojea and Loureiro (2011), the sensitivity to scope detected in this study is not restricted to CV and nor does it apply only to changes in a specific environmental good. This was further tested by interacting the scope dummy with the types of externalities and the valuation methods. For the interaction terms that resulted in a sufficient number of positive observations for valid analysis, this did not generate any significant effects, and is therefore not shown here. Although we are able to provide evidence for sensitivity to scope, we could not address the adequacy 2.5. Results 31 of scope sensitivity, i.e. whether the magnitude of response to a change in scope is appropriate. This is still a rather unresolved issue in scope sensitivity research (Desvousges, Mathews, and Train, 2012). The evidence for the impact of methodological variables on economic values is somewhat mixed. A clear result is provided by the coefficient for DCEs, indicating that DCEs result, ceteris paribus, in a higher economic value than CV and TCM. This finding is supported by some of the empirical evidence on the differences between values obtained by DCEs and CV (e.g. Hanley, Wright, and Adamowicz, 1998; Ryan and Watson, 2009). From a discounted utility point of view, assuming that the future has at least an infinitesimal weight, one would expect short-term payment durations to have a positive effect on economic values compared with long-term payments (Samuelson, 1937). However, the dummy for short-term payment durations does not turn out to be significant in any of the models. These results clearly indicate insensitivity to payment duration. A sensitivity analysis shows that this result remains robust when, in addition to payments of up to one year duration, payments that are limited to five and ten years are also defined as short-term payment durations, and only infinite payments are treated as long-term payments. This result may be interpreted in various ways. It might be that individuals have extremely high discount factors, and future costs therefore do not have an impact on their utility, even when these costs occur in the immediate future. However, even considering the high discount rates that are usually observed in economic experiments (for example, Harrison, Lau, and Melonie (2000) report annual discount rates close to 30%), it is still difficult to fully explain the non-significance of this variable. An alternative explanation could be that respondents simply do not consider longer payment durations during the surveys, and therefore show insensitivity to this factor. In contrast to the findings in Section 2.5.1, the models show that neither Asian nor American respondents attach significantly different values to hydropower externalities than European respondents, once control is included for other in- fluencing factors. The share of hydropower in the countries where studies were carried out does not seem to influence the economic values associated with hydropower externalities in the majority of models, although this variable is significant in Model 3. The dummy for users is positive and highly significant in all regression models, indicating that survey respondents who are direct users of the good that is affected by hydropower operation (mainly anglers who value the benefit of higher water flow of a river) are willing to pay significantly more 32 Chapter 2. Hydropower Externalities: A Meta-Analysis than other respondents to mitigate the effects on a resource (or to avoid its deterioration). Furthermore, the variable indicating the year when the survey was conducted (set to 0 for the earliest survey in 1983) is significant and positive in the three models, which suggests a significant time trend of increased economic values for the resources affected by hydropower over the years. This may be due to the growing scarcity of environmental goods or an increasing awareness in more recent years about the impacts of hydropower production. The dummy for income levels above the median is only significant in Model 3. This result points to a low income elasticity of demand with respect to hydropower externalities. Various combinations of random-effects have been tested in the mixed-effect Model 3. However, most of the variances of the random terms did not turn out to be significant. Likelihood-ratio tests indicated that the allocation of these terms to the random-effect part does not improve the model specification in most cases. Hence, the majority of variables are specified as fixed-effect terms. Only the inclusion of the dummy for the mitigation of the negative impacts on wildlife as a random term results in a significant model improvement. The variance of the variable is more than three times as large as its standard error. This suggests that there are significant differences with respect to the slope of the wildlife variable between studies. In other words, there are significant differences of the impact of valuing wildlife on welfare estimates between the studies, although the fixed- effect term of the same variable is not significant.

TABLE 2.4: Meta-analysis regression models

Model 1: Model 2: Model 3: WLS Random-effects Mixed-effects Variables Coeff. s.e. Coeff. s.e. Coeff. s.e. Constant 0.689 1.137 1.981 1.470 0.304 0.951 Type of externality and size of change valued Landscape & Vegetation 1.123 0.821 0.489 0.394 0.832** 0.331 (mitigation) Landscape & Vegetation -3.057*** 0.562 -3.454*** 0.359 -3.606*** 0.363 & Wildlife (deterioration) Wildlife (mitigation) -0.273 0.763 -0.038 0.483 0.205 0.749 Greenhouse gas -0.658 1.072 -1.489 0.906 -0.037 0.401 (improvement) Greenhouse gas 0.049** 0.024 0.075*** 0.023 0.028*** 0.010 (imp.) Hydropower ⇤ share 2.5. Results 33

Model 1: Model 2: Model 3: WLS Random-effects Mixed-effects Variables Coeff. s.e. Coeff. s.e. Coeff. s.e.

Aesthetics (mitigation) 0.257 0.979 1.116 0.984 -0.454 0.391 Aesthetics (deterioration) -0.282 0.780 1.019 0.853 -0.855** 0.353 Recreation (mitigation) -0.413 1.122 0.211 0.501 -0.285 0.522 Size of change (small) -0.788 0.500 -0.761*** 0.280 -0.670* 0.365 Methodological variables Valuation method (DCE) 3.234*** 0.851 3.193*** 0.843 3.598*** 0.690 Survey mode 0.886 0.836 0.597 0.722 1.135*** 0.422 (face-to-face surveys) Payment vehicle (tax 1.709 1.040 1.433 0.957 0.923 1.019 increase) Payment duration -1.315 0.835 -0.936 0.934 -0.203 0.74 (short-term) Sample characteristics North and South 1.204 0.821 1.064 0.972 0.318 0.623 America Asia -2.393 1.451 -1.418 1.280 -0.880 0.654 Hydropower share 0.010 0.013 0.004 0.013 0.014** 0.007 Users 6.556*** 1.364 6.561*** 0.957 6.174*** 0.885 High income (>median) 0.006 0.593 0.248 0.440 1.009*** 0.258 Year of study 0.150*** 0.046 0.104** 0.051 0.132*** 0.026 Random-effects (group variable: studies) 2 wildlife (mitigation) 5.387*** 1.736 2 constant 1.869*** 0.634 3.6e-25 2 residual 0.635*** 0.175 0.500*** 0.135 Model characteristics Log-likelihood -124.743 -114.922 AIC 293.485 273.844 BIC 346.163 326.522 R2 (Pseudo-R2) 0.770 (0.358) (0.409) Number of observations 81 81 81 Note: ***p<0.01, **p<0.05, *p<0.1.

2.5.3 Cross-validation

In order to compare Models 1, 2 and 3 and test for over-ﬁtting of the data, a cross- validation procedure was carried out. Cross-validation is a statistical technique similar to bootstrapping and jackkniﬁng but serving a different purpose. The main purpose of cross-validation is to obtain estimators of a model’s prediction 34 Chapter 2. Hydropower Externalities: A Meta-Analysis error, and compare the predictive power of various models (Efron and Gong, 1983). This procedure consists of several steps. First, 80% of the data points are randomly selected (the training set). Each model is then estimated based on the training set. Next, the values for the dependent variable of the remaining 20% of the data (the testing set) are predicted. The predicted values are compared with the actual values, and a standard error of the prediction is calculated. Formally, the prediction error has the following form:

1 N Prediction error = (y yˆ )2, (2.6) p v i i N u i uX t where yˆi denotes the predicted economic values that are compared with the actual values yi; and N is the number of observations included in the testing set. The procedure described above was repeated 10,000 times for all three models, resulting in a distribution of the prediction errors as depicted in Figure 2.1. 1 1.5 1.5 .8 1 1 .6 .4 .5 .5 .2 0 0 0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5 Model 1: Standard error of prediction Model 2: Standard error of prediction Model 3: Standard error of prediction

FIGURE 2.1: Histograms of the standard errors of model predic- tions based on 10,000 iterations

The mean value of the transformed dependent variable equals 7.01. The mean standard error of the prediction of Model 1 is 2.05, which is substantially reduced in Models 2 and 3 to 1.06 and 1.04, respectively. As expected, a panel specification substantially improves the predictive power of the model. Allow- ing between-study slopes of the wildlife mitigation regressor to vary results in a further reduction of the prediction errors, although the difference with Model 2 is small. The results of this cross-validation procedure provide evidence that Models 2 and 3 perform the best and Model 1 the worst out of the three model specifications. Further evidence confirming the superiority of Models 2 and 3 is obtained by simulating the expected error when applying the estimated meta-regression model for benefit transfer purposes. This is done by estimating the model based 2.6. Conclusions and discussions 35 on n-1 observations, and predicting the observation that is left out (e.g. Brander, Brouwer, and Wagtendonk, 2013). Comparing the predicted and the actually observed value, a prediction error can be calculated. This is then averaged across all observations. For the first model, this error amounts to 24%. On average, Model 1 applied to another context would thus result in an error margin of 24%. This error is considerably reduced in Model 2 (13%) and Model 3 (12%). Com- pared with the benefit transfer errors found in the literature, the simulated values obtained in this analysis are promising (e.g. Brouwer, 2000; Rosenberger and Stanley, 2006). However, the standard errors of this measure of prediction error have similar magnitudes as the prediction errors themselves.

2.6 Conclusions and discussions

This paper has applied meta-analysis, and estimated a meta-regression model in order to identify the factors that explain the variation in welfare estimates for the positive and negative external effects of hydropower production and to test for possible sensitivity to scope. The results revealed that welfare estimates for the external effects of hydropower are dependent on the type of externality assessed, as well as on whether deterioration or mitigation and improvements are valued. There is strong evidence for public aversion towards deteriorations in landscape, vegetation, and wildlife caused by hydropower. On the other hand, mitigation of the effects on these resources do not affect welfare measures significantly in most of the estimated models. The benefits of avoided greenhouse gas emissions are only significant in combination with the national share of hydropower in energy production. Sensitivity to scope is detected across externalities and valuation methods. The insights provided by this study are of considerable relevance for policies which aim to reduce the negative externalities of existing hydropower facilities, and for the planning processes of prospective hydropower plants. The importance of negative externalities and the lack of significant economic values for mitigating such effects constitute a rather unfavorable result for the future development and expansion of hydropower. This suggests the need for a strong public focus on the negative effects of hydropower, and a very limited willingness-to- pay for avoiding such effects. Hydropower projects in areas where the potential for negative externalities is high (e.g. in conservation areas) are therefore likely to meet with public resistance. Instead, hydropower plants will have to be planned in areas where they have as little impact as possible on the surrounding 36 Chapter 2. Hydropower Externalities: A Meta-Analysis landscape, vegetation, and wildlife. Claiming public financial resources for mitigating the effects of hydropower on environmental assets is hard to justify, in view of the fact that public willingness-to-pay for offsetting these externalities is so low. Furthermore, an expansion of hydropower has a higher chance of success when the positive externalities of avoiding greenhouse gas emissions are suffi- ciently large to compensate for the negative externalities of the energy source. This is more likely to be the case in countries with an already high share of hydropower in electricity production. Presumably, the populations in these countries have a higher level of awareness regarding the expected consequences of hydropower on greenhouse gas emissions. Only in those cases can the positive externality of hydropower production outweigh its negative effects. However, we showed that the relative change of the share of hydropower has to be in the order of at least 60 percentage points to compensate for the negative externalities, and there are not many countries in the world which could achieve such an expansion. Finally, aesthetic considerations do not seem to play an important role for the successful expansion of hydropower. This is in contrast to the key factors that drive public acceptance of other renewable sources of energy, in particular wind turbines. The visual effects of wind turbines have been identified as the key determinant of the public acceptance of wind power (e.g. Devine-Wright, 2005; Johansson and Laike, 2007; Warren et al., 2005; Wolsink, 2000). Although the literature on the factors which determine the acceptance of solar power is more limited, there is evidence that aesthetic considerations are also important for the case of photovoltaics (e.g. Faiers and Neame, 2006). This depends on whether photovoltaic structures are installed on existing artifacts, in which case they are not perceived as negative from an aesthetic point of view (Helena et al., 2015). Hence, it seems that each renewable energy source may have its own idiosyncratic factors that need to be considered in expansion planning processes, and what may be crucial for the development of one source of electricity may not be relevant for another. The average values obtained in our analysis seem generally applicable for benefit transfer purposes in cost-benefit analyses involving hydropower projects for a number of reasons. First, we find sensitivity to scope that is not limited to specific externalities or valuation methods. Secondly, the economic values do not differ significantly neither between different types of hydropower plants 2.6. Conclusions and discussions 37 nor between already existing and hypothetical new facilities. Finally, the prediction and transfer errors of our models are relatively low compared with those reported in the existing benefits transfer literature. Having said that, the general applicability of the results found in this study may be limited due to other factors that are likely to play a role in the probability of successfully expanding hydropower. Such factors include the topographical characteristics of regions where hydropower projects are planned, and the remaining share of free-flowing rivers in a country. Owing to data limitations, we could not control for either of these variables in our models. Furthermore, the non-representative country selection in our database is an issue to keep in mind. Specifically, developing countries are underrepresented, with China and Chile contributing only two studies and 13 observations to the data set. These two countries are considered as developing economies by the International Monetary Fund (IMF, 2014). Moreover, no low-income countries are included. Neverthe- less, we were able to control for continent of study origin which did not have a significant influence in any model specification. Finally, there are a number of methodological issues that need to be taken into account when interpreting the outcomes of this research. First of all, there is considerable heterogeneity with respect to the effects measured between the observations. Although an extensive number of independent variables were included in order to control for variations between the studies, and the explanatory and predictive power of the models is relatively high, it cannot conclusively be ruled out that there may be other factors that drive the valuation results. Further- more, the number of observations in the meta-regression is low, which is often the case in meta-analysis research. The trade-off between the conceptual homogeneity of the data studied and the amount of data points available for analysis is a general issue in meta-analysis research. The relative scarcity of studies on the effects of hydropower and their valuation also shows that this is a rather underinvestigated area that calls for further research. 38 Chapter 2. Hydropower Externalities: A Meta-Analysis

2.A Studies included in the meta-analysis

# Study 1 Loomis, J., Sorg, C., & Donnelly, D. (1986). "Economic Losses to Recreational Fisheries due to Small-head Hydro-power Development: a Case Study of the Henry’s Fork in Idaho". Journal of Environmental Management 22 (1), pp. 85–94. 2 Kosz, M. (1996). "Valuing riverside wetlands: the case of the "Donau-Auen" national park". Ecological Economics 16, pp. 109–127. 3 Navrud, S. (1995). Hydro Fuel Cycle. Part II (p.127-249) in European Commission DG XII Science Research and Innovation (1995): ExternE: Externalities of Energy. Volume 6: Wind and Hydro. EUR 16525 EN, European Comission Publishing. Luxembourg. Navrud, S. (2001). "Environmental costs of hydro compared with other energy options". Hydropower and Dams 8 (2), pp. 44–48. 4 Biro, Y. E. K. (1998). "Valuation of the Environmental Impacts of the Kayraktepe Dam/Hydroelectric Project, Turkey: An Exercise in Contingent Valuation". Ambio 27 (3), pp. 224–229. 5 Loomis, J. (1996). "Measuring the economic benefits of removing dams and restoring the Elwha River: Results of a contingent valuation survey". Water Resources Research 32 (2), pp. 441–447. 6 Hansesveen, H., & Helgas, G. (1997). "Environmental Costs of Hydropower Development - Estimering av miljokostnader ved en vannkraftutbygging i Ovre Otta". Norwegian University of Life Sciences, As, Norway. 7 Bergland, O. (1998). Valuing Aesthetical Values of Weirs in Watercourses with Hydroelectric Plants - Verdsetjing av estetiske verdiar i tilknytning til tersklar i regulerte vassdrag. Oslo: Norwegian Water Resources and Energy Directorate (NVE). 8 Filippini, M., Buchli, L., & Banfi, S. (2003). "Estimating the benefits of low flow alleviation in rivers: the case of the Ticino River". Applied Economics 35, pp. 585–590. 9 Hanley, N., & Nevin, C. (1999). "Appraising renewable energy developments in remote communities: the case of the North Assynt Estate, Scotland". Energy Policy 27 (9), pp. 527–547. 10 Loomis, J. (2002). "Quantifying recreation use values from removing dams and restoring free-flowing rivers: A contingent behavior travel cost demand model for the Lower Snake River". Water Resources Research 38 (6), pp. 2–1–2–8. 11 Han, S.-Y., Kwak, S.-J., & Yoo, S.-H. (2008). "Valuing environmental impacts of large dam construction in Korea: An application of choice experiments". Environmental Impact Assessment Review 28 (4-5), pp. 256–266. 12 Sundqvist, T. (2002). "Power Generation Choice in the Presence of Environmental Externalities". PhD Thesis, Lulea University of Technology, Lulea, Sweden. Retrieved from https://pure.ltu.se/portal/files/153854/LTU-DT-0226-SE.pdf 13 Bothe, D. (2003). Environmental Costs due to the Karahnjukar Hydro Power Project on Iceland. University of Cologne: Department of Economic and Social Geography, Cologne, Germany. 14 Hynes, S., & Hanley, N. (2006). "Preservation versus development on Irish rivers: whitewater kayaking and hydro-power in Ireland". Land Use Policy 23 (2), pp. 170–180. 15 Bergmann, A., Colombo, S., & Hanley, N. (2008). "Rural versus urban preferences for renewable energy developments". Ecological Economics 65, pp. 616–625. 2.A. Studies included in the meta-analysis 39

# Study

16 Håkansson, C. (2009). "Costs and benefits of improving wild salmon passage in a regulated river". Journal of Environmental Planning and Management 52 (3), pp. 345–363. 17 Navrud, S. (2004). Environmental Costs of Hydropower, Second Stage - Miljøkostnadsprosjektet Trinn 2. EBL report 181. 18 Longo, A., Markandya, A., & Petrucci, M. (2008). "The internalization of externalities in the production of electricity: Willingness to pay for the attributes of a policy for renewable energy". Ecological Economics 67 (1), pp. 140–152. 19 Kataria, M. (2009). "Willingness to pay for environmental improvements in hydropower regulated rivers". Energy Economics 31 (1), pp. 69–76. 20 Robbins, J. L., & Lewis, L. Y. (2009). "Demolish it and they will come: Estimating the economic impacts of restoring a recreational fishery". Journal of the American Water Resources Association 44 (6), pp. 1488–1499. 21 Ku, S.-J., & Yoo, S.-H. (2010). "Willingness to pay for renewable energy investment in Korea: A choice experiment study". Renewable and Sustainable Energy Reviews 14 (8), pp. 2196–2201. 22 Aravena, C., Hutchinson, W. G., & Longo, A. (2012). "Environmental pricing of externalities from different sources of electricity generation in Chile". Energy Economics 34 (4), pp. 1214–1225. 23 Ponce, R. D., Vasquez, F., Stehr, A., Debels, P., & Orihuela, C. (2011). "Estimating the Economic Value of Landscape Losses Due to Flooding by Hydropower Plants in the Chilean Patagonia". Water Resources Management, 25(10), pp. 2449–2466. 24 Kosenius, A.-K., & Ollikainen, M. (2013). "Valuation of environmental and societal trade-offs of renewable energy sources". Energy Policy, 62, pp. 1148–1156. 25 Ehrlich, Ü., & Reimann, M. (2010). "Hydropower versus Non-market Values of Nature: A Contingent Valuation Study of Jägala Waterfalls, Estonia". International Journal of Geology, 4(3), pp. 59-63 26 Guo, X., Liu, H., Mao, X., Jin, J., Chen, D., & Cheng, S. (2014). "Willingness to pay for renewable electricity: A contingent valuation study in Beijing, China". Energy Policy 68, pp. 340–347. 27 Gogniat, S. (2011). "Estimating the Benefits of an Improvement in Water Quality and Flow Regulation: Case study of the Doubs". Master’s Thesis, Université de Neuchâtel, Neuchâtel, Switzerland. Retrieved from https://www.unine.ch/files/live/sites/iaf/files/shared/documents/ S%C3%A9minaires%20en%20finance/S%C3%A9minaires%20%C3%A9conomie% 20et%20finance/SG_Thesis_final.pdf 28 Klinglmair, A., & Bliem, M. (2013). Die Erschliessung vorhandener Wasserkraftpotenziale in Österreich im Spannungsfeld von Energiepolitik und ökologischen Schutzzielen. 8. Internationale Energiewirtschaftstagung an der TU Wien. Retrieved from https://www.researchgate.net/publication/261223345_Die_ Erschliessung_vorhandener_Wasserkraftpotenziale_in_Osterreich_im_ Spannungsfeld_von_Energiepolitik_und_okologischen_Schutzzielen 29 Klinglmair, A., Bliem, M., & Brouwer, R. (2012). Public Preferences for Urban and Rural Hydropower Projects in Styria using a Choice Experiment. IHS Kärnten Working Paper. Retrieved from http://www.carinthia.ihs.ac.at/HydroVal/files/working_paper.pdf

Chapter 3

Choice Certainty, Consistency, and Monotonicity in Discrete Choice Experiments

3.1 Introduction

This study investigates choice certainty, choice consistency, and choice monotonicity in the context of discrete choice experiments (DCE). There is heterogeneity in the choice literature on the definition of these terms. Our definitions follow the most common practice in the DCE literature, as identified by a literature review of studies which assess the determinants of certainty, consistency, and monotonicity (see Section 3.2). "Choice certainty" henceforth refers to "stated choice certainty", as it is based on respondents’ self-reported answers to a question on choice certainty after each choice task. In the broader microeconomic literature, "choice consistency" is often linked to the axiom of transitivity (e.g. Jehle and Reny, 2001). However, we follow the definition by Carlsson, Mørk- bak, and Olsen (2012), and define choice consistency as "making the same choice in two equal choice tasks". An identical definition of choice consistency is used in the DCE literature on the determinants of consistency in a test-retest setting (Brouwer, Logar, and Sheremet, 2016; Mørkbak and Olsen, 2015; Rigby, Burton, and Pluske, 2016; Schaafsma et al., 2014). Finally, "choice monotonicity" is assumed to hold if a respondent chooses a non-dominated alternative in a choice task that contains a dominated alternative, which is a hypothetical alternative that is worse than at least one other alternative in a choice task with respect to all attributes. 42 Chapter 3. Choice Certainty, Consistency, and Monotonicity

The rationale to investigate these concepts simultaneously is threefold. First, in contrast to the existing literature, an integrated analysis of the three constructs on the basis of an identical sample of respondents allows us to examine how they relate to each other and to identify their common and idiosyncratic drivers. We thereby specifically focus on two measures of choice-task complexity. Second, although a number of studies investigate potential determinants of choice certainty and choice consistency separately, we are not aware of any studies in the field of environmental economics that aim to identify the drivers of choice monotonicity in DCEs. As well as such a novel investigation, the simultaneous analysis allows us to assess whether the drivers of choice monotonicity are common to the determinants of choice certainty and consistency, or whether they are idiosyncratic. Finally, the existence of all three concepts disagrees with the conventional assumptions made in microeconomics that people know their preferences perfectly, and that these preferences are stable, coherent, and rationally maxi- mized (e.g. Brown et al., 2008; Rabin, 1998). Choice monotonicity in the framework of microeconomics is related to the axiom of consumer choice, which, in its strict formulation, implies that more is always better than less (e.g. Jehle and Reny, 2001). Accordingly, in the context of a DCE, a choice alternative should be preferred over another alternative when the first is better on one attribute and at least as good on all the other attributes. The existence of choice uncertainty, inconsistency, and non-monotonicity is hence typically ruled out in DCEs. How- ever, even if evidence for their existence is found, this does not necessarily imply choice irrationality. Preferences are, for example, allowed to shift between choice tasks (Rigby, Burton, and Pluske, 2016). In addition, a key assumption in random utility theory (McFadden, 1974), which underlies the DCE method, is the existence of an informational asymmetry between the decision maker and the analyst. Thus, it is not always straightforward to identify all the underlying causes of choice behavior from the choice information collected in a DCE. Choice behavior may be, for example, driven by attributes which are present in a respondent’s utility function but not included in a DCE. In this case, a behavioral violation of conventional assumptions would reflect more a misspecification of the DCE than evidence for choice irrationality. It is nevertheless relevant to detect the degree to which choice uncertainty, inconsistency, and non-monotonicity are present, since they affect the outcomes of choice models. Moreover, identifying their determinants is of interest, as it may help to minimize their causes. The occurrence of uncertainty, inconsistency, and non-monotonicity is especially likely in DCEs, since respondents are often confronted with unfamiliar goods and/or attributes. 3.1. Introduction 43

Even if the characteristics of the good are familiar to a respondent, it may still be cognitively challenging to conceptualize specific attribute levels (Rigby, Burton, and Pluske, 2016). In the first part of this paper, we aim to compare choice models based on choice responses that differ with respect to choice certainty, consistency, and monotonicity. We first investigate stated choice certainty. The information on choice certainty at the choice-task level enables us to investigate possible dy- namic effects of choice certainty over the course of the experiment. Furthermore, we assess the effect of including choice certainty questions in the survey. For this purpose, we use a split-sample approach and present the choice certainty questions to only half of the respondents. The procedural effect of posing a question on certainty has, to the best of our knowledge, only been studied in Brouwer et al. (2010). We add to this study by also investigating the potential effects caused by the presence of certainty questions at a choice-task-specific level. Next, in order to analyze choice consistency, the first choice task is randomly repeated at varying positions in the choice-task sequence within the same DCE. This allows us to detect the degree of choice consistency, and to check for locational effects of the repeated task on consistency. Finally, choice monotonicity is assessed by including a choice task that involves a dominated choice alternative in a choice set. The second part of the analysis comprises logit models that regress choice certainty, consistency, and monotonicity on possible determinants in order to identify common and unique drivers of these concepts. In doing so, we focus on investigating which of two measures of choice-task complexity, i.e. entropy or the utility difference between the chosen and the second-best alternative in a choice task, has a higher explanatory power in the models. This is a novel contribution to the literature, since these two measures of choice-task complexity have not been compared before. The remainder of this paper is structured as follows: Section 3.2 provides a literature review on choice certainty, consistency, and monotonicity. The literature review aims to identify the key determinants and hypotheses that will be used in the subsequent logit models. Section 3.3 describes the econometric models, and Section 3.4 the case study and the DCE design. Section 3.5 presents the results, and there follows a discussion and conclusions in Section 3.6. 44 Chapter 3. Choice Certainty, Consistency, and Monotonicity

3.2 Choice certainty, consistency, and monotonicity in discrete choice experiments

The literature on choice certainty in the context of DCEs can be divided into two streams, one of which investigates how to account for choice certainty in choice models, while the other focuses on the determinants of choice certainty. Within the first stream of literature, there are studies that consider stated choice certainty in choice models by excluding, recoding, or weighting responses according to their self-reported choice certainty (e.g. Beck, Fifer, and Rose, 2016; Beck, Rose, and Hensher, 2013; Kosenius, 2009; Lundhede et al., 2009). Since treating choice certainty as an explanatory construct in this manner raises the issue of endogeneity, some authors have switched to more complex models which treat stated choice certainty as a latent construct (e.g. Dekker et al., 2016). Other studies focus on inferred rather than on stated choice certainty by assuming that a higher choice certainty is reflected in a lower error variance (Brouwer et al., 2010; Uggeldahl et al., 2016). The second stream of research consists of studies that estimate mainly logit or probit models in order to investigate the determinants of stated choice certainty. These studies, and the significant determinants they detect, are listed in Table 3.1. The most common definition of choice consistency in the DCE literature is to make identical choices when faced with identical choice tasks1. Therefore, the usual test for choice consistency is to repeat one or several choice tasks using either the same or different samples of respondents. This is implemented between separate choice survey occasions (test-retest) which implies a time-lag (Bliem, Getzner, and Rodiga-Laßnig, 2012; Brouwer, Logar, and Sheremet, 2016; Liebe, Meyerhoff, and Hartje, 2012; Mørkbak and Olsen, 2015; Rigby, Burton, and Pluske, 2016; Schaafsma et al., 2014) or within the same survey which results in almost no time-lag or a lag in the order of few minutes (Brouwer et al., 2010; Brown et al., 2008; Campbell, 2007; Carlsson, Mørkbak, and Olsen, 2012; Hanley, Wright, and Koop, 2002; Rigby, Burton, and Pluske, 2016; Rulleau and Dachary-Bernard, 2012; Scarpa, Campbell, and Hutchinson, 2007; Soliño et al., 2012). There is, again, a stream of literature that investigates the drivers of choice consistency in DCEs by estimating binomial models, listed in Table 3.2. Note that all studies except Carlsson, Mørkbak, and Olsen (2012) investigate choice consistency in a test-retest setting.

1Some studies test for alternative deﬁnitions of choice consistency. Brown et al. (2008), for example, focus on circular triads, that is, on preferences which imply A>B>C>A. 3.2. Choice certainty, consistency, and monotonicity in DCEs 45

TABLE 3.1: Studies that regress stated choice certainty on its determinants

Authors (year of Model estimated Signiﬁcant determinants (direction of the publication) effect)a Brouwer et al. (2010) Ordered probit Age (+) Credibility of the alternatives (+) Experience with the DCE’s topic (+) Familiarity with the information provided (+) Income (+) Male gender (+) Utility differenceb (+) Dekker et al. (2016) Latent variable Credibility of the alternatives (+) model approach Male gender (+) Entropy (-) Survey length (-) Hensher, Rose, and Generalized Differences between the attribute levels of the Beck (2012)c ordered logit status quo and non-status quo alternatives (+) Environmental attitude (+) Stated number of acceptable alternatives per choice set (+) Income (-) Number of children in household (-) Kosenius (2009) Binary logit Acceptance of scenario elements (+) Male gender (+) Positive attitudes towards the scenarios (+) Time spent at the resource valued (+) University degree (+) Utility differenced (+) Olsen et al. (2011)e Ordered probit Income (+) Utility differenceb (+) Uggeldahl et al. (2016) Fixed effects Utility difference (+) regression model Choice of the status quo option (+) Gaze shifts (eye-tracking) (-) Later position of the task (-) Notes: aSome studies also included attributes and scenario-related variables as regressors. These are not listed. bBetween the chosen and the second-best alternative. cCertainty in terms of how certain it is that respondents will choose in a real situation the same alternative as they indicated they would choose in the choice tasks. dBetween the alternatives with the highest and the lowest utility in a choice task. eThe results of two case studies are reported. We only report the effects that proved to be stable across both case studies. 46 Chapter 3. Choice Certainty, Consistency, and Monotonicity

TABLE 3.2: Studies that regress choice consistency on its determinants

Authors (year of Model estimated Signiﬁcant determinants (direction of the publication) effect) Brouwer et al. (2016) Binary probit Choice certainty (stated) (+) Credibility of the information provided (+) Differences between the choice task attribute levels (+) Membership in an environmental organization (+) Carlsson, Mørkbak, Binary probit Choice of the status quo option in the ﬁrst and Olsen (2012) sequence (+) Later position of the repeated task (+) Stated certainty that the chosen alternative provides the largest utility (+) Utility differencea (+) Mørkbak and Olsen Binary probit Utility differencea (+) (2015) Income (-) Rigby, Burton, and Binary probit Willingness to take on more complex tasks (+) Pluske (2016) Entropy (-) Schaafsma et al. (2014) Binary probit None Note: aBetween the chosen and the second-best alternative.

A considerable amount of literature in health economics includes a test for monotonicity in the form of a choice task, where one alternative is superior to the other alternatives on all choice attributes (e.g. De Bekker-Grob, Ryan, and Gerard, 2012; Determann et al., 2017; Lancsar and Louviere, 2006; Özdemir et al., 2010; Soliño et al., 2012). Based on the outcome of choice monotonicity tests, a part of the survey participants are typically excluded from further analysis in health studies. Examples of DCEs in environmental economics that include a choice task with a dominant or dominated alternative are: Bateman, Munro, and Poe (2008), Campbell (2007), Foster and Mourato (2002), Hanley, Wright, and Koop (2002), and Scarpa, Campbell, and Hutchinson (2007). However, the literature that investigates the determinants of choice monotonicity is limited and, to the best of our knowledge, non-existent in environmental economics. In health economics, San Miguel, Ryan, and Amaya-Amaya (2005), for example, test for monotonicity as suggested in this study and, in addition, run further monotonicity tests and regress their results on explanatory variables. Their most important finding implies that respondents who state they have great difficulty with a choice task are less likely to pass the monotonicity test, i.e. to choose the theoretically expected alternative. They also detect a positive effect of higher 3.2. Choice certainty, consistency, and monotonicity in DCEs 47 education on choice monotonicity, and a bell-shaped effect of age, indicating that both younger and older respondents have a lower probability of passing the monotonicity test. Finally, an inverted bell-shaped effect of the repeated task position reflects the situation that choice consistency is determined by learning effects for an early position of the repeated choice task, and by fatigue setting in for later positions (Swait and Adamowicz, 2001a). The review of the literature reveals several key determinants of choice certainty, consistency, and monotonicity. The most important regressors identified are included in our analysis in Section 3.5.3. Variables that proved to be significant in at least three studies include gender, where we expect male respondents to state a higher choice certainty than female respondents. Furthermore, income seems to be important, although the direction of the effect is unclear: Brouwer et al. (2010) report a positive effect on choice certainty, whereas Hensher, Rose, and Beck (2012) find the opposite. Mørkbak and Olsen (2015) confirm the latter with respect to choice consistency. The authors argue that this may be due to a lower importance of the price attribute for high-income respondents, which adds more randomness to their choices. A further ambiguous variable is higher education, which is reported to have a positive effect on choice certainty and monotonicity by, respectively, Kosenius (2009) and San Miguel, Ryan, and Amaya-Amaya (2005). On the other hand, Olsen et al. (2011) and Dekker et al. (2016) control for it, but find no effect. Based on the literature review, we also expect that positive (environmental) attitudes positively affect choice certainty and consistency. Finally, there is ambiguity in the use of either entropy or utility difference as an indicator for choice-task complexity. Our analysis hence aims at comparing the explanatory power of the two measures. Both measures of task complexity have different origins: Entropy originates from information theory (Shannon, 1948; Soofi, 1994; Swait and Adamowicz, 2001b). In the realm of DCEs, entropy describes the equality of a choice task’s alternatives based on the predicted individual specific probabilities of choosing each alternative. We hypothesize that the higher the entropy, and therefore the more similar the choice alternatives in a choice task, the higher is the choice-task complexity. Higher choice-task complexity, in turn, is expected to reduce choice certainty, consistency, and monotonicity. Utility difference, in contrast, has its roots in Wang (1997). We follow the literature’s adaptation of Wang’s hypothesis to a DCE context, and hypothesize that the choice complexity of a choice task is highest when the utility difference between alternatives is lowest (Brouwer et al., 2010; Carlsson, Mørkbak, and 48 Chapter 3. Choice Certainty, Consistency, and Monotonicity

Olsen, 2012; Mørkbak and Olsen, 2015; Olsen et al., 2011). Hence, a higher utility difference (and therefore a presumably lower task complexity) is expected to increase choice certainty, consistency, and monotonicity. Finally, we include two paradata variables related to time: The time spent by survey respondents reading informational pages that appear before the actual choice tasks, and the time used per choice task. The latter has been shown to outperform eye-tracking information in predicting stated choice certainty (Uggeldahl et al., 2016). We expect the time used per choice task to have a negative impact on stated choice certainty, and the time used for reading informational pages to positively affect certainty, consistency, and monotonicity.

3.3 Econometric Models

The Swait-Louviere test procedure (Swait and Louviere, 1993) is applied to assess whether there are differences between the choice behavior of respondents who: (i) state they are certain or uncertain about their choices; (ii) choose consistently or inconsistently; (iii) were and were not asked a question on choice certainty; and (iv) were subjected to different positions of the repeated choice task in the choice sequence. The test procedure for verifying the equality of choice behavior between all these groups of respondents consists of two steps. In a ﬁrst step, the equality of preference parameters between different samples is assessed. For- mally, the following hypothesis is tested:

SLA H0 : 1 = 2 = pooled. (3.1)

This hypothesis is tested by estimating attribute-only mixed logit (MXL) models (see equation 3.5) for two different samples, as well as for a pooled sample. In the latter case, the scale parameters (µ1 and µ2) are allowed to differ between the samples. Formally, the hypothesis in equation 3.1 is assessed by the likelihood ratio (LR) test-statistic:

µ =µ LR = 2 LL 16 2 (LL + LL ) . (3.2) pooled 1 2 ⇥ ⇤ SLA If H0 is rejected, one cannot conclude whether the difference between the samples is caused by differences in preference parameters or differences in both preference and scale parameters, because they are confounded (Louviere, Hensher, SLA and Swait, 2000). If, however, H0 cannot be rejected, one is able to proceed to 3.3. Econometric Models 49 the second step of the test procedure and isolate differences in scale parameters between two samples: SLB H0 : µ1 = µ2 = µpooled. (3.3)

Note that scale parameters are inversely related to the error variance. Equation 3.3 is tested by comparing the log-likelihood of the pooled model which allows for different scale parameters between the samples and another pooled model that does not allow for the scale factors to vary:

µ =µ µ =µ LR = 2 LL 1 2 LL 16 2 . (3.4) pooled pooled ⇥ ⇤ If equation 3.4 is not rejected, we cannot reject the equality of scale parameters between the two samples. Logit models are used for regressing choice certainty, choice consistency, and choice monotonicity on their determinants. More specifically, we estimate binary logit models to identify the determinants of choice consistency and monotonicity, and ordered logit model to assess the drivers of choice certainty (see, e.g., Cameron and Trivedi (2005) for the econometric specifications of these models). The latter are random-effects panel specifications, since choice certainty is analyzed at a choice-task level, whereas choice consistency and monotonicity models can only be estimated at the respondent level. Amongst other regressors, two measures of choice-task complexity are included: The expected utility difference between the chosen alternative and the second-best alternative, as well as entropy. The procedure to calculate both of these measures begins with estimating an attribute-only MXL model based on the following MXL probability distribution for each individual n, alternative i, and choice task t:

x e 0n nit Pnit()= f()d. (3.5) nxnjt j e 0 Z ⇣ ⌘ P On the basis of equation 3.5, the MXL model parameters are estimated by applying maximum likelihood estimation on the appropriate log-likelihood function. This model does not include any sociodemographic covariates, since the logit models that follow already control for these. In order to determine the expected utility difference, the individual-speciﬁc aggregate utility for each choice alternative in each choice task is calculated based on the MXL model results. Finally, the difference in expected utility between the observed choice j and the second- best alternative (either k or l in our context with three choice alternatives (see Section 3.4.1)) for each choice task t and each individual n is computed. The 50 Chapter 3. Choice Certainty, Consistency, and Monotonicity utility difference can formally be speciﬁed as follows (e.g. Olsen et al., 2011):

Utility difference = E U (x , " ) nt njt njt njt max E U (x ," ) ; E U (x , " ) = (3.6) nkt nkt nkt nlt nlt nlt h i = ˆ0 x max ˆ0 x ; ˆ0 x . n njt n nkt n nlt ⇥ ⇤ A similar procedure was applied to calculate entropy again using the estimated MXL model. The entropy measure utilizes the predicted choice probabilities P obtained from the MXL model, and estimates a value of choice-task complexity for each individual n and each choice task t as follows (Rigby, Burton, and Pluske, 2016; Shannon, 1948; Sooﬁ, 1994; Swait and Adamowicz, 2001b):

I Entropy = P (i)lnP(i) 0. (3.7) nt i=1 X For the case of three choice alternatives in a choice task, entropy can reach a maximum value of approximately 1.1, given that the probability of choosing an alternative is identical for all three alternatives. In contrast, if the probability of choosing an alternative equals 1, entropy reaches its minimum value of 0.

3.4 Case-study description

3.4.1 Discrete choice experiment

Following the nuclear accident in Fukushima, Japan, the Swiss government released a policy paper on the long-term Swiss energy strategy in 2011 ("Energy Strategy 2050" (SFOE, 2013)). Our study incorporates two main strategic goals of this energy strategy: The expansion of electricity produced by hydropower, and the phase-out of nuclear power plants. Hydropower and nuclear power constitute by far the most important sources of electricity production in Switzerland, with respective shares in total electricity production of 60% and 34% in 2015 (SFOE, 2016). From a policy perspective, an expansion of hydropower renders a phase-out of nuclear power more likely, since the probability and the consequences of a loss in electricity production that such a phase-out may involve could be reduced. Due to this interlinkage of the envisaged policies for the two main sources of electricity in Switzerland, the preferences for hydropower and 3.4. Case-study description 51 nuclear power should correspondingly be studied simultaneously. For this reason, we designed a DCE that focuses on a hypothetical expansion of hydropower and takes into account the expected decrease in the provision of nuclear power. The DCE includes three choice alternatives: two unlabeled hydropower expansion scenarios, and one status quo scenario. The scenarios are characterized by four attributes and their associated levels, as summarized in Table 3.3.

TABLE 3.3: Attribute and attribute levels in the DCE

Attribute Attribute levels in hypothetical Attribute levels in alternatives status quo alternative Type of hydropower Extending existing hydropower No hydropower expansion plants expansion Construction of new hydropower plants

Lifetime risk of death 20% increase in risk (1 in 750,000 Current risk (1 in from a dam breach people are expected to die) 900,000 people are expected to die) 40% increase in risk (1 in 650,000 people are expected to die)

Lifetime risk of death 60% decrease in risk (1 in 7 million Current risk (1 in 3 from a nuclear accident people are expected to die) million people are expected to die) 30% decrease in risk (1 in 4 million people are expected to die)

Increase in household’s 100, 200, 300, 400, 500, 600 No change in the annual electricity bill annual electricity bill (CHF)

The first attribute describes whether the current hydropower plants would be extended, or whether new facilities would be constructed. It is explained to survey respondents that the construction of new plants is associated with stronger detrimental environmental effects compared with the extension of existing plants. The second and third attributes reflect the changes in the risk of death associated with an expansion of hydropower and a reduction in nuclear power production, respectively. Both risk attributes can take on three levels, one of which appears only in the status quo alternative. The absolute risk magnitudes of both attributes were carefully constructed on the basis of data from existing studies (Burgherr and Hirschberg, 2008, 2014; Hirschberg et al., 2016). The relative changes in risk are based on realistic scenarios and reflect the type of hydropower plants that would be built or extended (only plants with dams or various types of plants) and the number of nuclear power plants in Switzerland 52 Chapter 3. Choice Certainty, Consistency, and Monotonicity that would be phased-out2. Care was taken in communicating the risks of death to respondents with the help of risk ladders. The risk ladders put the risks associated with hydropower and nuclear power into context by comparing them with other risks of death. Finally, a price attribute was included which was framed as an increase in a household’s yearly electricity bill3. The results of two pretest rounds with a total of 570 respondents (see Section 3.4.3) was used to define and calibrate meaningful attribute levels. Figure 3.1 presents an example of a choice task. Each respondent received a sequence of eight choice tasks.

A) Expansion B) Expansion C) No expansion

Type of hydropower

expansion New construction Extension No expansion

Risk of dying from a

dam breach +40% risk +20% risk Current risk (1 in 650,000 people) (1 in 750,000 people) (1 in 900,000 people)

Risk of dying from a

-30% risk -60% risk nuclear accident Current risk (1 in 4,000,000 people) (1 in 7,000,000 people) (1 in 3,000,000 people)

Increase in your household‘s yearly +200 CHF/year +300 CHF/year +0 CHF/year electricity bill

FIGURE 3.1: Choice task example

3.4.2 Elicitation of choice certainty, consistency, and monotonicity

Choice certainty was elicited for every choice task. That is, we asked the survey participants the following question after each choice task: "How certain are

2We assume that, even for the case of a complete phase-out of nuclear power in Switzerland, French nuclear plants that are located within 40km of the Swiss border would still be a cause of risk. 3Note that the average annual electricity bill per household in Switzerland roughly equals 930 CHF (ca. 930 USD) (Elcom, 2014) 3.4. Case-study description 53 you about your choice?". The question appeared on the same page as the choice task. A Likert-type answer scale with the following five options was presented: not certain at all, not certain, neither certain nor uncertain, certain, very certain. We used a split-sample approach, so that one sample of respondents was asked about their choice certainty (sample 1), while the other sample did not receive this question (sample 2). All other elements of the survey were identical for the two samples. This allows us to investigate the procedural effect of asking respondents about their choice certainty. Choice consistency was elicited by re- peating the first choice task once again later in the choice-task sequence. The position of the repeated choice task varied between the 5th,6th,7th, and 8th (last) position in the choice-task sequence, and was randomly assigned to each respondent. Finally, we included a test for choice monotonicity by adding a choice task with a dominated alternative. In this choice task, one of the two hypothetical alternatives was worse than the other hypothetical alternative with respect to all attributes. Although this choice task includes a strongly dominated choice alternative, there is no unique dominant alternative identifiable, since the non- dominated choice alternative and the status quo alternative do not dominate each other. This is because in our DCE design the levels of some attributes in the status quo alternative are better and others are worse than their levels in the non-dominated choice alternative. The position of the choice task involving a dominated alternative was fixed at the 4th position in the choice-task sequence for all respondents.

3.4.3 Sampling procedure and choice experiment design

Three rounds of survey pretests were conducted in order to test the understandability of the survey and of the DCE, as well as to derive prior estimates needed for generating the final choice design. First, 20 respondents were asked to fill in a paper-and-pencil version of the survey, followed by a personal interview. Sec- ond, 220 respondents, who were representative for the German-speaking part of Switzerland, were recruited to answer a first online-version of the DCE. In a third pretest, another 350 respondents answered the online survey. The final survey version was completed in June 2016 with a sample consisting of 502 respondents. The response rate was 16.3%. The samples for pretests and for the final survey were recruited by Intervista AG, a market research company with a panel of 50,000 registered individuals throughout Switzerland. The final sample of 502 respondents consisted of two independent samples comprising 248 and 54 Chapter 3. Choice Certainty, Consistency, and Monotonicity

254 respondents. The only distinction between them is the presence/absence of the choice certainty questions. Both samples are representative for the German- and French-speaking population of Switzerland which accounts for roughly 95% of the total Swiss population. The representativeness was safeguarded by independent recruitment which ensured equality of the samples and the general population with respect to gender, age, education, and the geographical distribution of residence within Switzerland. The independent recruitment of samples is also expected to minimize the impact of unobserved respondent heterogeneity (Dekker, Koster, and Brouwer, 2014). Additional important features of the survey (Menegaki, Olsen, and Tsagarakis, 2016) include the following: Question- naire invitations were sent by email which contained links that directed respondents to the survey web page. The respondents were allowed to drop the survey and resume it at a later point. Duplication of answers by the same respondents was prevented by IP-tracking and by inspection of the sociodemographic information and the respondents’ paradata. Paradata collected include the time and date of initiating and completing the survey, the time used for each survey page, data on the characteristics of the device and the operating system which were used to answer the survey, IP-address, and approximate information on the lon- gitudes and latitudes of the respondents’ location. At the end of the survey, the respondents were asked a few follow-up questions about the survey itself, which revealed that it was well understood. Finally, those respondents who completed the survey received widely redeemable gift vouchers equivalent to 4 CHF. We generated a D-efficient DCE design in Ngene 1.1.2. Bayesian priors that are bounded by the results of the second and third pretests, and that follow a normal distribution were used for the design-generating process. Such a specification allows for a random instead of a fixed distribution of priors, and hence takes the approximate character of pretest priors into account. A status quo option was included in the design generation using an alternative-specific constant (ASC) for the status quo alternative. Dominant alternatives were excluded in the design generation procedure, but a choice task which involved a dominated alternative was added to the choice sequence after the design generation process. The final experimental design consisted of two blocks of choice sets, each of which contained eight choice tasks. The two blocks were randomly assigned to the survey participants. 3.5. Results 55

3.5 Results

3.5.1 Descriptive results

Following the criteria of Bateman et al. (2002), we identiﬁed 31 protest responses which were excluded from further analysis. This leaves us with a remaining sample population of 235 respondents for sample 1 (which includes a question on choice certainty) and 248 respondents for sample 2 (without a question on choice certainty). No statistically signiﬁcant differences between the samples with respect to any of the variables that were controlled for at the point of re- cruiting could be found. This holds even after excluding protest respondents (Table 3.4).

TABLE 3.4: Comparison of samples 1 and 2 with respect to sociodemographic characteristics

Variable Type Description Test Test- p- Statistic value Income Ordinal 15 income Mann-Whitney 0.772 0.440 categories Gender Bivariate 0=Male; Chi-squared 0.376 0.540 1=Female Age Continuous Respondents’ Mann-Whitney 0.069 0.945 age (15-84) Language Bivariate 0=German; Chi-squared 0.746 0.389 spoken 1=French University Bivariate 0=No university Chi-squared 0.106 0.744 degree degree; (Bachelor’s, 1=University Master’s, PhD) degree Note: ***p<0.01, **p<0.05, *p<0.1.

Figure 3.2 reports the choice-task-specific responses to the question on choice certainty for sample 1. The distribution of certainty values is left-skewed. On average, 64.4% of respondents stated that they were certain or very certain about all choice tasks. The dynamics of stated choice certainty across the choice tasks is limited: The share of respondents who indicated that they were certain or very certain about their choice slightly increases after the first choice task, but decreases again after the fourth task. The share of respondents who stated that they are either not certain or not certain at all decreases from 19.6% in the first choice task to 14.9% in the last choice task. Note that the 4th choice task includes 56 Chapter 3. Choice Certainty, Consistency, and Monotonicity a dominated alternative. As expected, the share of certain and very certain (not certain at all and not certain) respondents is highest (lowest) for this task. Sta- tistical tests confirm that there is limited dynamics across the choice tasks: The Kruskal-Wallis test statistic for equality of choice certainty across all tasks cannot be rejected at the 10% significance level. The Mann-Whitney test for pairwise comparisons of choice certainty between the tasks rejects equality in 79% of all possible comparisons, and only finds significant differences between the tasks 1 and 4, 1 and 8, 2 and 4, 2 and 8, 4 and 5, and 5 and 8. Finally, the variance of stated choice certainty is lower within than between respondents, since almost 47% of the respondents indicated the same level of choice certainty across all choice tasks.

50 not certain at all 40 not certain 30 neither certain

Shares(in %) 20 nor uncertain certain 10 very certain 0 1st 2nd 3rd 4th 5th 6th 7th 8th Choice task

FIGURE 3.2: Stated choice certainty across choice tasks

The inferred choice consistency according to the corresponding position of the repeated choice task is illustrated in Figure 3.3. Two insights can be gained: First, the respondents in sample 2 seem to be more consistent than the respondents in sample 1, pointing to a possible interaction effect between the presence of a choice certainty question and choice consistency. Second, the choice consistency of respondents who were subjected to the repeated choice task later in the choice sequence seems to be slightly lower, which is especially true for sample 1. Nevertheless, none of these effects is statistically signiﬁcant. The Mann-Whitney test for the equality in choice consistency between the two samples cannot be rejected (p-value of 0.14). Within-sample equality between the respondents who differ with respect to the position of the repeated choice task cannot be rejected 3.5. Results 57 either, when based on the Kruskal-Wallis test procedure (p-value of 0.67 and 0.98 for, respectively, samples 1 and 2).

Consistent 65 respondents of sample 1 60 Consistent

Shares(in %) respondents 55 of sample 2

50 5th 6th 7th 8th Position of the repeated task

FIGURE 3.3: Choice consistency according to the position of the repeated choice task

The share of monotonic respondents who did not choose the dominated alternative in the choice task that includes such an alternative are similar between the samples, with 88.1% and 86.7% of respondents choosing monotonically in, respectively, sample 1 and 2. These values are higher than the shares of consistent respondents. While choice monotonicity is higher for sample 1, choice consistency is higher for sample 2. The differences in monotonicity and consistency between samples 1 and 2 are, however, not statistically signiﬁcant.

3.5.2 Swait-Louviere test results

The Swait-Louviere test procedure as described in Section 3.3, was run to test the equality of choice behavior between: (i) respondents who were certain and uncertain about their choices; (ii) consistent and inconsistent respondents; (iii) samples 1 and 2; and (iv), respondents who were subjected to the repeated choice task at different positions in the choice-task sequence. The test results are shown in Tables 5 (i-ii), 6 (iii), and 7 (iv). Table 3.5 reports the results of the Swait-Louviere tests that compare the choice models estimated separately for certain and uncertain respondents, as well as for consistent and inconsistent respondents. The test for equality of 58 Chapter 3. Choice Certainty, Consistency, and Monotonicity choice models estimated for monotonic and non-monotonic respondents could not be performed because the number of respondents categorized as non-monotonic was too low. Certain respondents were defined as respondents who stated that they were either certain or very certain about their choices. Survey participants who stated that they were neither certain nor uncertain (i.e. who selected the response category "neither certain nor uncertain") are hence classified as uncertain. This definition results in 46% certain and 54% uncertain respondents. The first step of the Swait-Louviere test procedure convincingly rejects the equality of certain and uncertain, as well as the equality of consistent and inconsistent respondents (column 5 in Table 3.5). In this case, it is meaningless to proceed with the second step of testing the equality of scale parameters, which for this reason is not reported in Table 3.5. In order to compare choice behavior between samples 1 and 2, Swait-Louviere tests were applied at the sample level and at the individual choice-task level. In the latter case, two choice tasks were merged for each test to ensure a sufficient number of observations for the identification of the models. That is, the choice tasks 1 and 2, 3 and 4, 5 and 6, and 7 and 8 were merged and, as such, compared for the two samples. Two main conclusions can be derived from the results presented in Table 3.6. First, the hypothesis of equality of preference parameters between the samples 1 and 2 (row 1, column 5), as well as the test for equality of scale parameters (row 1, column 6), cannot be rejected. In other words, the choice behavior between the respondents in samples 1 and 2 is not significantly different. This result implies that asking the respondents about their choice certainty does not lead to systematic differences in their choices. Second, the test for equality of the samples with respect to specific (pairs of) choice tasks confirms this result, since the equality of preference and scale parameters between the samples cannot be rejected for any task combinations. This means that the presence of a choice certainty question also has no effect on choices while controlling for the position of choice tasks within the sequence. The results of the Swait-Louviere tests on the equality of choice behavior between samples that differ with respect to the position of the repeated choice task show that the equality of the preference parameters cannot be rejected for any of the comparisons (column 5 in Table 3.7). In other words, different positions of the repeated choice task are not associated with procedural biases. This con- form to our expectations, since respondents were not informed either about the presence of a repeated task or about its position. Neither can the majority of the tests for the equality of the scale parameters reject equality (column 6 in Table 3.5. Results 59

3.7). Significant differences in scale are found between the samples that received the repeated task at the 5th and 6th position, and weakly significant differences between the samples that were shown the repeated task at the 6th and 7th position. The results concerning differences in scale parameters suggest that the error variance of the sample with a repeated choice task at the 5th position is generally higher than the error variance of samples that were shown the repeated task at a later position (column 8 in Table 3.7). A possible explanation is that the 5th choice task appears immediately after the task involving a dominated alternative, which could have caused increased randomness in the following choice task. 60 Chapter 3. Choice Certainty, Consistency, and Monotonicity ) ) 2 2 2 2 Rel. Rel. / / 2 2 1 1 ( ( variance variance ) ) 2 2 /µ /µ 1 1 µ ( µ ( Rel. scale Rel. scale sample is fixed at 1.) d st d n/a n/a n/a n/a n/a n/a (1df) -value p (1df) LR-test: -value p LR-test: nd 2 nd 2 c c -value (11df) p LR-test: -value (11df) st p LR-test: 1 st 1 b ) b 1 ) µ 1 µ = = 1 LL pooled µ 1 ( LL pooled µ ( a ) a ) 1 1 µ µ 6= 6= 1 1 µ LL pooled ( µ LL pooled ( nd nd LL 2 LL 2 sample sample inconsistent (sample 1 and 2 merged) respondents st st (1) (2) (3) (4) (5) (6) (7) (8) (1) (2) (3) (4) (5) (6) (7) (8) LL 1 sample sample <0.1; LL: Log-likelihood; df: degrees of freedom. <0.1; LL: Log-likelihood; df: degrees of freedom. 3.6: Swait-Louviere test results for equality of choice behavior between samples 1 and 2 p p ABLE T <0.05, * <0.05, * 3.5: Swait-Louviere test results for certain vs. uncertain respondents (sample 1 only) and consistent vs. p p ABLE T <0.01, ** <0.01, ** Test for differences in the scale parameters between the samples. Pooled MXL model constraining the scale parameters to be equal between the samples. Pooled MXL model allowing the scale parameters to vary between the samples (the scale parameter of the 1 Test for differences in the preference parameters between the samples. p p a b c d Tasks 1 & 2 - sampleTasks 3 1 & vs. 4 2 - sampleTasks 5 1 & vs. 6 2 - sampleTasks 7 1 & vs. 8 2 - sample 1Notes: vs. *** -424.821 2 -361.595 -427.809 -462.513 -399.478 -382.052 -437.435 -893.229 -407.668 -747.710 -871.809 -893.255 -811.260 -748.311 -871.813 -812.477 0.380 0.702 0.285 0.693 0.819 0.273 0.927 0.119 1.04 0.76 1.02 0.69 0.92 1.73 0.96 2.10 Choice samplesAll tasks - sample 1 vs. 2 LL -1408.279 1 -1380.541 -2794.920 -2795.189 0.349 0.464 0.94 1.13 Certain vs. uncertainNotes: *** -826.174 -557.462 -1407.469 -1408.279 0.00*** Respondents’ choice characteristics Consistent vs. inconsistentMonotonic vs. non-monotonic not enough respondents to estimate -1594.651 a model based on non-monotonic respondents only -1115.740 -2767.317 -2795.189 0.00*** 3.5. Results 61 ) 2 2 / 2 1 Rel. variance ( ) 2 /µ 1 µ Rel. scale ( d LR-test: -value nd 2 p (1df) c LR-test: -value st 1 p (11df) b ) 1 µ = 1 µ LL pooled ( a ) 1 µ 6= 1 µ LL pooled ( tion of the repeated choice task (samples 1 and 2 merged) nd LL 2 sample <0.1; LL: Log-likelihood; df: degrees of freedom. p st <0.05, * 3.7: Swait-Louviere test results for equality of choice behavior between samples that differ in the posi- p (1)LL 1 sample (2) (3) (4) (5) (6) (7) (8) ABLE T <0.01, ** p Position of the repeated choice task 5th vs. 6th5th vs. 7th5th vs. 8th -666.0556th vs. 7th -666.0556th vs. 8th -666.0557th vs. 8th -828.457 -828.457Notes: *** -828.457 -659.276 -659.276 -629.770 -659.276 -1495.979 -629.770 -1327.242 -629.770 -1300.589 -1495.979 -1488.794 -1327.242 -1462.262 -1300.589 -1292.725 0.992 -1488.794 0.975 -1462.262 0.573 -1292.725 0.998 0.707 0.016** 0.769 0.292 0.177 0.075* 0.74 0.255 0.536 0.86 0.84 1.24 1.16 1.83 0.92 1.35 1.42 0.65 0.74 1.18 62 Chapter 3. Choice Certainty, Consistency, and Monotonicity

3.5.3 Logit model results

The results of the logit models are shown in Table 3.8. The models on choice certainty are estimated on sample 1 only, because only the respondents in this sample received the choice certainty questions. The models on choice consistency and monotonicity are estimated on the pooled sample of respondents. The effect of the follow-up certainty questions on choice consistency and monotonicity is controlled for by a dummy variable for sample 1. Apart from theoretical expectations about the determinants of choice certainty, consistency, and monotonicity based on the existing literature, log-likelihood ratio tests were applied to find the model with the best fit. Two models are presented for each concept in Table 3.8. The two model specifications differ in the measure of complexity of the choice task: Entropy serves as an explanatory variable in the first set of models (Models 1, 3, and 5), and utility difference in the second set of models (Models 2, 4, and 6). The two measures of complexity are included in separate regressions since they are highly correlated (the correlation coefficient across choice tasks and respondents is equal to -0.62). This procedure allows us to compare their relative impact on the dependent variables, while avoiding confounding caused by multicollinearity. The correlations of the remaining variables in each regression are reasonable, with only a few pairs of variables that have correlation coefficients which are larger than 0.2. Excluding these variables does not affect the main conclusions. Note that the ordinal variable for choice certainty was reduced from five to four levels by merging the "not certain at all" and "not certain" response categories, because there were only a few observations in the most uncertain category. No case study-specific attributes are included in the analysis in order to ensure the generalizability of the results. Furthermore, choice task-specific effects are controlled for by choice-task-related variables, e.g. the variables which reflect the position of a choice task in the choice sequence and the complexity measures. McFadden’s (1974) pseudo-R2 values suggest that the models explain a rather small part of the variation in the data. However, we consider this not to be critical for three reasons: First, many of the pseudo-R2 values found in the literature, if they are reported at all, are of comparable magnitudes (e.g. Colombo, Glenk, and Rocamora-Montiel, 2016; Hensher, Rose, and Beck, 2012; Kosenius, 2009; Mørkbak and Olsen, 2015); second, pseudo-R2 values are, in general, considerably lower than the traditional R2 values obtained in OLS regressions (McFadden, 1979); and, third, the pseudo-R2 is of minor interest in the context of this paper, as we are more interested in testing relationships than in building prediction models. 3.5. Results 63

TABLE 3.8: Logit regression results

Choice certaintya Choice consistencyb Choice monotonicityb (1) (2) (3) (4) (5) (6) Variable Coeff. Coeff. Coeff. Coeff. Coeff. Coeff. (s.e.) (s.e.) (s.e.) (s.e.) (s.e.) (s.e.)

Choice-task-related variables Position of choice task -0.240* -0.225* in sequence (1 to 8) (0.131) (0.131) Position of choice task 0.025* 0.024* in sequence (squared) (0.013) (0.013) Time spent per choice -0.733*** -0.745*** task (log) (0.141) (0.141) Entropy -3.628*** -0.965 -3.766 (0.540) (1.267) (3.349) Utility difference 0.216*** 0.220* 1.123*** (0.060) (0.119) (0.301) Question on choice certainty -0.247 -0.233 -0.314 -0.323 (1=yes, 0=no) (0.204) (0.204) (0.318) (0.327) Position of the repeated task -0.305 -0.316 (1=7./8., 0=5./6.) (0.204) (0.204) Individual-speciﬁc variables Income above sample 0.375 0.370 -0.024 -0.024 0.935** 0.945** average (1=yes, 0=no) (0.512) (0.506) (0.214) (0.215) (0.369) (0.376) Female (1=yes, 0=no) -1.936*** -1.924*** 0.380* 0.330 1.011*** 0.944*** (0.501) (0.496) (0.212) (0.214) (0.350) (0.361) Age -0.002 -0.001 0.010 0.010 -0.032*** -0.033*** (0.014) (0.014) (0.006) (0.006) (0.010) (0.010) University education 0.530 0.533 0.564** 0.587** 0.028 0.133 (1=yes, 0=no) (0.532) (0.526) (0.230) (0.230) (0.372) (0.382) Recreates in/at waters 1.612* 1.609* 0.073 0.056 1.568*** 1.737*** (1=yes, 0=no) (0.945) (0.934) (0.388) (0.390) (0.475) (0.494) Member of an env. org. 0.147 0.181 0.700*** 0.752*** 0.783** 0.959*** (1=yes, 0=no) (0.502) (0.496) (0.215) (0.217) (0.359) (0.365) Survey 2.577*** 2.594*** 0.292 0.313 0.425 0.553 understandabilityc (0.596) (0.589) (0.237) (0.238) (0.363) (0.371) Time spent reading 0.092 0.093 0.024 0.027 0.070 0.089* informational pages (0.059) (0.059) (0.025) (0.025) (0.049) (0.051) Model characteristics Constant 0.385 -0.928 3.616 -3.039** (1.377) (0.566) (2.581) (1.260) d µ1 -6.553*** -2.718** (1.449) (1.345) d µ2 -3.929*** -0.147 (1.440) (1.342) 64 Chapter 3. Choice Certainty, Consistency, and Monotonicity

Choice certaintya Choice consistencyb Choice monotonicityb (1) (2) (3) (4) (5) (6) Variable Coeff. Coeff. Coeff. Coeff. Coeff. Coeff. (s.e.) (s.e.) (s.e.) (s.e.) (s.e.) (s.e.)

d µ3 2.145 5.826*** (1.435) (1.354) 12.392*** 12.113*** (1.589) (1.552) # Observations 1,880 1,880 470 470 470 470 # Respondents 235 235 470 470 470 470 Log-likelihood -1402.9 -1419.7 -282.9 -281.4 -137.9 -129.9 McFadden pseudo R2 0.047 0.035 0.055 0.060 0.157 0.205 Notes: ***p<0.01, **p<0.05, *p<0.1. aRandom-effects ordered logistic regression. bBinomial logistic regression. c1=very/rather, 0=partly/rather not/not at all. dCut-off points of the ordered logistic regression.

The results of the models on stated choice certainty indicate that the position of the choice task in the choice sequence has an inverted bell-shaped effect on choice certainty. This supports the results of San Miguel, Ryan, and Amaya- Amaya (2005), but, in our case, with respect to choice certainty. Since the choice certainty analysis embraces all choice tasks faced by each respondent, we include the time used per choice task as a regressor. This proves to exert a highly significant and negative effect, meaning that survey participants who used more time to complete a choice task are more likely to report a low choice certainty for that task. On the other hand, the time a respondent spent on reading relevant informational pages of the survey has no statistically significant effect on choice certainty. However, the coefficients almost reach the 10% significance level with p-values of 0.12 and 0.11 in the Models 1 and 2, respectively. All these informational pages (seven in total) appeared before the choice tasks, and conveyed information on hydropower, nuclear power, their associated risks, the expansion scenarios of hydropower, and the DCE. None of the informational pages contained survey questions. Comparing the two models for choice certainty that differ in the measure of choice-task complexity reveals that both measures contribute in explaining stated choice certainty, but the model that includes entropy has a slightly better model fit. As expected, the coefficient of entropy is negative, implying that higher entropy, associated with increased choice-task complexity, results in lower choice certainty. The only sociodemographic variable that turns out to have a significant effect on choice certainty is gender. As expected, female 3.5. Results 65 respondents seem to be less certain about their choices than male respondents. In contrast with the existing literature, we cannot report any effect of income on choice certainty. Respondents who state the survey to be very or rather understandable are significantly more likely to have a higher choice certainty. Fi- nally, the random effects parameter, , is significant, meaning that stated choice certainty is correlated across the choice tasks within the set of an individual’s responses (Olsen et al., 2011). Log-likelihood ratio tests that compare Models 1 and 2 with ordered logit models without panel specifications show that there exists sufficient variability between individuals, and hence that a panel specification performs better than a standard ordered logistic regression4. Turning next to the models which explain choice consistency, the binary variable indicating university education becomes significant, suggesting that respondents with a university education (Bachelor’s, Master’s or PhD degree) are more consistent in their choices than those with a lower education level. In contrast to our expectations based on the results from the existing literature, education is a unique determinant of choice consistency, and does not drive choice certainty and monotonicity. This may suggest that choice consistency is, to a higher degree than choice certainty and monotonicity, influenced by an individual’s cognitive abilities. Female respondents are more consistent in their choice, although the variable on gender does not quite reach the 10% significance level in model 4 (p-value of 0.12). Note that we find the opposite effect of gender in the case of choice certainty, which shows that female respondents are less certain about their choices. A further factor which increases the likelihood of making consistent choices is membership of an environmental organization, which is likely to serve as an indicator of environmental attitudes. This confirms the findings in the literature with respect to choice consistency. Ultimately, only the utility difference, but not the entropy of the repeated choice task when shown for the first time, appears to influence choice consistency. Our results on the determinants of choice monotonicity suggest that female respondents are not only more likely to choose consistently but are also more monotonic. Age and income uniquely affect choice monotonicity: Income increases choice monotonicity while younger survey respondents have a higher chance of choosing monotonically than older respondents. However, the square root of age has no impact in any of the six models. We hence cannot confirm the bell-shaped effect of age which is reported in San Miguel, Ryan, and Amaya- Amaya (2005). As other variables are unaffected by the inclusion of the square

4The results are available from the authors on request. 66 Chapter 3. Choice Certainty, Consistency, and Monotonicity root of age, we decided to exclude this variable in the final models. As in the case of choice consistency, the utility difference of the choice task which involves a dominated alternative appears significant when included in the regression, while entropy remains insignificant. A larger utility difference is associated with a lower choice-task complexity, so, as expected, its impact on choice monotonicity is positive.

3.6 Discussion and conclusions

This study simultaneously investigates choice certainty, choice consistency, and choice monotonicity in the context of DCEs. Several conclusions can be drawn from the results of the Swait-Louviere tests. First, choice behavior significantly differs between certain and uncertain, as well as between consistent and inconsistent, respondents. Second, we do not find a procedural effect caused by the inclusion of a choice certainty question after each choice task on certainty, consistency, or monotonicity. The procedural equivalence with respect to choice certainty confirms the findings of Brouwer et al. (2010). In addition to Brouwer et al. (2010), we also investigate whether the same result holds at a choice-task level, and find that it does. In other words, there exists neither an overall nor a sequence-dependent procedural effect of asking a question on choice certainty. Third, as expected, the equality of choice behavior between samples that differ by the position of the repeated choice task cannot be rejected, which implies that the position of a repeated choice task in a choice-task sequence does not lead to any systematic changes in choice behavior. We identify a number of idiosyncratic determinants for each of the three concepts. However, we only find limited evidence for common drivers of choice certainty, consistency, and monotonicity, suggesting that these are rather separate constructs. The factors that prove to have an effect on all three concepts are the utility difference of choice tasks and gender, although the evidence for the impact of gender on choice consistency is rather weak. Female participants are less certain about their choices, confirming the findings of Brouwer et al. (2010), Dekker et al. (2016), and Olsen et al. (2011). At the same time, female respondents more often choose monotonically and consistently. Comparing the two measures of choice-task complexity, we are able to report that both measures are relevant for choice certainty, whereas only the utility difference is important for explaining choice consistency and monotonicity. The importance of the utility 3.6. Discussion and conclusions 67 difference for explaining choice certainty and consistency is in line with the literature. Rigby, Burton, and Pluske (2016) report that choice consistency decreases with higher complexity of choice tasks as measured by entropy, whereas we find no such effect. However, the comparability of their finding with our study may be limited because Rigby, Burton, and Pluske (2016) examine choice consistency in a test-retest setting, whereas we focus on consistency within a choice-task sequence. We are not aware of any previous study that tests for the influence of entropy or utility difference on choice monotonicity. The implications of our main findings for the DCE literature are manifold. Significant differences between the choice behavior of certain and uncertain and consistent and inconsistent survey participants mean that it is necessary to account for these factors in choice models. The drivers of choice certainty, consistency, and monotonicity identified in this study could be used as explanatory variables in choice models in order to achieve this goal and possibly mitigate the issue of endogeneity. In concrete terms, there is evidence that it is necessary to ensure the utmost cognitive ease of a survey, as several results show that this is an important factor in reducing choice uncertainty, inconsistency, and non-monotonicity. A university education and the understandability of the survey, both of which are also likely to be associated with the cognitive ease of responding to the choice tasks, are expected to increase choice consistency and certainty. The time spent on choice tasks has a negative effect on choice certainty, and the time spent reading information pages positively affects choice monotonicity. This suggests that a strict division of the survey into an informational part before the actual choice tasks, and self-explanatory choice tasks which do not necessitate additional clarification, is required. Procedurally, no reserva- tions against directly asking survey respondents about their choice certainty can be confirmed. We find that the utility difference as a measure of task complexity is a superior explanatory factor of choice consistency and monotonicity than entropy. In terms of choice experiment design, this finding suggests that choice- task complexity needs to be limited as much as possible. Put differently, it is recommended to avoid choice tasks that involve alternatives which are expected to result in very similar utility values. This accords with the advice of Hensher, Rose, and Greene (2015), who suggest not using too narrow ranges of attribute levels in designing DCE. Further research combining the analysis of choice certainty, consistency, and monotonicity could provide additional evidence on common and idiosyncratic drivers of these concepts. Although the results in Table 3.8 provide insights into 68 Chapter 3. Choice Certainty, Consistency, and Monotonicity the common and idiosyncratic drivers of choice certainty, consistency, and monotonicity, they do not answer the question of whether they also influence each other. In order to explore this issue, we estimated separate models on the three concepts that include each other as regressors, while acknowledging that these models are likely to suffer from endogeneity (the models are not presented here, but are available from the authors on request). Including choice certainty, consistency, or monotonicity as regressors did, apart from introducing collinearity, usually not turn out to have a significant impact in the models. This provides some evidence that there is limited interlinkage between these concepts. An instrumental variable (IV) approach would, however, allow this question to be answered more reliably. Since we did not find instruments that exclusively describe any of the three concepts with sufficient predictive power, and, at the same time, are not expected to be correlated with the other concepts, we were not able to conduct such an analysis. If other unique determinants were found, or those identified in this study were confirmed by other studies that analyze all three concepts simultaneously, it may be justified to estimate instrumental variable models which can provide further evidence for (the lack of) possible interrela- tions between the three concepts. Finally, the results of this study suggest strong effects of gender on choice certainty, consistency, and monotonicity. Although a rather extensive literature on gender differences exists in other areas of economics, such as behavioral experiments (e.g. Charness and Gneezy, 2012) or marketing (e.g. Meyers-Levy and Loken, 2015), there are only a few valuation studies in environmental economics which explicitly focus on gender effects. An example is Ladenburg and Olsen (2008), who report a starting-point bias induced by an instructional choice set for female but not for male respondents. The authors explain their findings by the "selective hypothesis" of Meyers-Levy (Meyers-Levy, 1989; Meyers-Levy and Loken, 2015), which states that females have a tendency to process information more comprehensively than males, who are more selective information proces- sors and more often rely on heuristics. Testing the selective hypothesis and, more generally, an in-depth examination of the gender effects in the realm of DCE con- stitutes another area for further research. 69

Chapter 4

A Comparison of Attribute non-Attendance in Discrete Choice Experiments based on Stated, Inferred, and Mouse-Tracking Data1

4.1 Introduction

Attribute non-attendance (ANA) describes a behavioral phenomenon in discrete choice experiments. It refers to the situation where choice experiment participants ignore one or several attributes in their decision process. ANA can be explained from an economic, experimental design, or psychological perspective. From an economic point of view, the key assumption underlying choice experiments is that people make trade-offs between all the good characteristics presented to them in the form of choice attributes (Colombo, Christie, and Hanley, 2013; Scarpa et al., 2009). In other words, indifference curves are assumed to be continuous, which follows from the axiom of continuous preference relations (Arrow and Debreu, 1954; Debreu, 1959). Individuals who ignore one or more

1This chapter has been submitted to a journal for review. It was first presented at the 23rd Annual Conference of the European Association of Environmental and Resource Economists in Athens in June 2017. 70 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments choice attributes in their choice process have, for example, lexicographic preferences which cannot be represented by a continuous utility function (e.g. Jehle and Reny, 2001). From an economic perspective, ANA thus represents incom- plete preference structures. An alternative explanation is related to the design of stated choice experiments, specifically the range of choice attribute levels that is presented to respondents (Hensher, Rose, and Greene, 2012). According to this reasoning, ANA does not necessarily mean that a respondent is unwilling to trade off the non-attended attribute in a real choice setting, but only over the attribute level range that is presented in a choice experiment. If this is the case, ANA observed in an experiment is not actual ANA, but rather an artifact resulting from a misspecification of attribute level ranges in the discrete choice experiment. Finally, there exists a psychological foundation to ANA that goes beyond the explanations found in economic theory or experimental design. The psychological perspective has its roots in Kahneman and Tversky’s dual-phase model of decision-making (Kahneman and Tversky, 1979). This model assumes that the decision process consists of two stages. In the first stage, the decision problem is edited, whereas in the second stage the edited problem is evaluated, and a decision is made. The editing stage serves to rearrange the decision concerned such that the subsequent decision-making is simplified. In the context of discrete choice experiments, the editing stage may consist of a number of strategies that facilitate the choice of an alternative. Such decision strategies, or heuristics, include, for example, aggregation of similar attributes, cancellation of, or non- attendance to, an attribute, and the referencing of an attribute around one’s own experience (Hess and Hensher, 2010; Hess, Rose, and Hensher, 2008). From this point of view, ANA is a component of the editing stage of the decision process that helps to simplify the subsequent choice. Independently of the perspective taken to explain ANA, its existence under- mines the fundamental idea that discrete choice experiments are rooted in the ability and willingness of respondents to make trade-offs between different attributes. Failing to consider ANA in a choice analysis leads to biased parameter estimates which, in turn, result in biased welfare estimates (e.g. Carlsson, Kataria, and Lampi, 2010). Pooling observations based on participants considering all choice attributes, with observations based on respondents who show non-compensatory behavior in trade-offs, leads to biased and erroneous statistical estimation (Scarpa et al., 2009). It is, therefore, crucial to take ANA into account when estimating choice models from choice experiment data. 4.2. Attribute non-attendance 71

In this paper, we compare a novel visual ANA approach based on mouse- tracking data with the two most common approaches for addressing ANA, i.e. stated and inferred ANA. Visual ANA refers to collecting information on ANA by tracking the visual information acquisition behavior of respondents in a discrete choice experiment. The scarce literature on visual ANA uses eye-tracking in a laboratory setting. In contrast, this is the ﬁrst study that investigates ANA in discrete choice experiments using mouse-tracking. Contrary to eye-tracking, mouse-tracking makes it possible to track visual behavior online, and therefore takes advantage of important web-based survey features such as reduced logisti- cal survey costs and access to a larger number of respondents. The main objective of this study is to examine how these alternative approaches for measuring ANA compare with each other in terms of model performance, and whether the visual information obtained by mouse-tracking is better suited to explain ANA behavior than the standard approaches based on stated or inferred ANA information. Since stated ANA is a subjective measure of ANA and inferred ANA an indirect choice data driven measure of ANA, we hypothesize that the direct visual information obtained through mouse-tracking provides a more reliable measure of ANA behavior, and thus results in a superior model ﬁt. Finally, this study compares the performance of models that use mouse-tracking information on ANA with those reported in the few studies that use eye-tracking to investigate visual ANA. The remainder of this paper is structured as follows. Section 4.2 summarizes the existing literature on ANA. Section 4.3 describes the case study and choice experiment design and Section 4.4 the elicitation of stated and visual ANA. Sec- tion 4.5 explains the choice modeling approach. The results are presented in Section 4.6. Finally, in Section 4.7, the results are discussed and conclusions are drawn.

4.2 Attribute non-attendance

In the discrete choice experiment literature, ANA has been most widely researched in transport economics and, slightly less extensively, in environmental and health economics. The most common approaches for investigating ANA in the existing literature include stated and inferred ANA. The former approach collects information about ANA from survey participants, while the latter approach captures ANA with econometric modeling techniques. A third approach is visual 72 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments

ANA, which is still infrequently used and so far includes only studies using eye- tracking. Studies on stated ANA usually directly ask respondents in a survey whether they have considered all attributes in their choices and, if not, which ones they have ignored. This can be asked either as a binary question (whether an attribute has been attended to or not) or in terms of the strength of (non)-attendance to an attribute. Stated ANA, if asked after the choice task sequence, measures serial ANA, i.e. ANA over an entire sequence of choice tasks. Most of the literature on stated ANA focuses on serial stated ANA, because there is no convincing theoretical reason for ANA to be choice-task specific (Balcombe, Fraser, and McSor- ley, 2015). Moreover, asking for ANA after each choice task poses an additional cognitive challenge to survey participants, and may lead to fatigue. There are only a few studies focusing on choice-task specific ANA (e.g. Scarpa, Thiene, and Hensher, 2010). Within the choice experiment literature using stated ANA, there are further differences with respect to how the information on stated ANA is incorporated into econometric models. The most common approach is to restrict the coefficients of choice attributes to zero for respondents stating non- attendance to these attributes (e.g. Alemu et al., 2013; Carlsson, Kataria, and Lampi, 2010; Hensher, Rose, and Bertoia, 2007; Hensher, Rose, and Greene, 2005; Kehlbacher, Balcombe, and Bennett, 2013; Kragt, 2013; Puckett and Hensher, 2008, 2009). Alternatively, two separate coefficients for each choice attribute can be estimated for respondents attending or not attending to an attribute. This approach has been applied in Colombo, Christie, and Hanley (2013), Hess and Hensher (2010) and Scarpa et al. (2013). Finally, Balcombe, Burton, and Rigby (2011), Campbell, Hutchinson, and Scarpa (2008), Carlsson, Kataria, and Lampi (2010), and Nguyen et al. (2015) include the information on stated ANA in choice models in the form of interaction terms between each choice attribute and a binary variable that equals one if a respondent stated non-attendance to an attribute and zero otherwise. The estimated main effect in this case represents the marginal value of an attribute for respondents who attend to the attribute, whereas the sum of main and interaction effects, which is expected to be close to zero, represents the marginal value for respondents who do not attend to the attribute. Although the elicitation of stated ANA information and its use in choice models is common in the literature, there is a lack of trust in the validity and reliability of stated ANA responses for a number of reasons (e.g. Balcombe, Burton, and Rigby, 2011; Hensher and Greene, 2010; Hensher and Rose, 2009; Weller et al., 4.2. Attribute non-attendance 73

2014). One of them is the belief that, instead of ignoring an attribute, respondents actually only assign a low importance to it. There might be cases where respondents state non-attendance to an attribute, but in fact still have a positive, albeit small, marginal utility for this attribute (Hess and Hensher, 2010). Fur- thermore, conditioning the econometric analysis of choice experiment data on stated ANA information obtained after the choice experiment could result in endogeneity (e.g. Hess and Hensher, 2013). Stated ANA responses may correlate with unobserved variables, which also influence the stated choices, implying that they are not independent from the error component of the choice model (Collins, Rose, and Hensher, 2013). A final reservation about using stated ANA information in estimating choice models is that one reason for the analysis of ANA is a kind of mistrust in the respondents’ choice processing capabilities. If there is doubt about the mental capabilities of choice experiment participants to trade-off all the attributes presented in a choice situation and hence ANA is expected, it is not entirely clear why researchers should trust the stated information about non-attendance by the same participants. For this reason, part of the literature has explored other ways of analyzing ANA, most prominently by inferring information on ANA from the choice experiment data. In this context, latent class (LC) models are the most widely used choice models. In contrast to the standard applications of LC models in discrete choice analysis, the different latent classes are defined beforehand, and have a behavioral meaning that represents specific non-attendance patterns (Hensher, Rose, and Greene, 2015). More specifically, the classes in LC models used for analyzing ANA reflect which attributes (and combinations thereof) are not attended to. The parameters for these attributes are restricted to zero. The estimated class membership probabilities represent the share of respondents who behave according to a specific non-attendance strategy. Most frequently, the equality-constrained LC model (ECLC) is applied. In this model the parameters associated with the attended attributes are restricted to be the same across the different classes. This approach ensures that the different classes only differ in terms of their behavioral meaning representing (non)-attendance to different attributes. Studies that use this approach include Hensher and Greene (2010), Hole (2011), Kragt (2013), Lagarde (2010), and Scarpa et al. (2009). Campbell, Hensher, and Scarpa (2011) and Hensher and Greene (2010) present models which allow parameters for attended attributes to vary across classes. Independently of whether parameters are restricted to be equal across the classes or not, these models do not allow parameters to vary across respondents 74 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments within a class. Hess et al. (2013) argue that this might lead to a possible confounding between ANA and taste heterogeneity. Assuming that all respondents within a certain class, which describes an ANA behavioral rule, do not attend to an attribute ignores the fact that some respondents within the class may actually attend to the attribute, but they attach only a small marginal value to it. For these respondents, a low sensitivity to an attribute is mistakenly identified as non-attendance to this attribute. To remedy this issue, Hess et al. (2013) suggest a framework that combines an LC model with a continuous mixed logit model. In this combined model, the parameter for the ANA class is still restricted to zero, but the parameters within the ANA classes are specified as random terms and are hence allowed to vary across respondents. Similar models are used by Collins, Rose, and Hensher (2013), Hole, Kolstad, and Gyrd-Hansen (2013), Hensher, Collins, and Greene (2013), and Hess et al. (2013). Whereas Collins, Rose, and Hensher (2013), Hess et al. (2013), and Hole, Kolstad, and Gyrd-Hansen (2013) find an improvement in LC models when random parameters are introduced, the improvement is minor in the case of Hensher, Collins, and Greene (2013). These studies are, however, not directly comparable, since Hensher, Collins, and Greene (2013) account for attribute processing rules other than ANA within the same model, and they do not restrict the fixed and random parameters to be equal across the classes unlike the three other studies. Nevertheless, a common finding in the literature is that the share of respondents displaying ANA behavior tends to reduce after the introduction of random parameters (Hensher, Collins, and Greene, 2013; Hess et al., 2013; Hole, Kolstad, and Gyrd-Hansen, 2013). This suggests that ANA and taste heterogeneity may be confounded, and that LC models with fixed parameters may assign some degree of taste heterogeneity to ANA, leading to the overestimation of ANA behavior. Hensher, Rose, and Greene (2015) conduct an extensive study which estimates four models with different latent classes (including different numbers of ANA and full attribute attendance (FAA) classes), as well as fixed and random parameters within the classes. The results of this study are unable to confirm the superiority of models allowing for random effects in a LC framework. They show only a very small improvement in model performance compared with models which have fixed parameters. While there is no agreement in the literature on which choice model specification performs best for the analysis of inferred ANA, there seems to be more consensus with respect to the better model fit of inferred over stated ANA models (Campbell, Hensher, and Scarpa, 2011; Kragt, 2013; Scarpa et al., 2013; Weller et al., 2014). 4.2. Attribute non-attendance 75

Scarpa et al. (2013) suggest using more sophisticated methods based on tracking respondents’ attribute information processing, in order to collect more valid information about ANA. The most obvious technology for tracking information processing is eye-tracking. Eye-tracking has been used widely in marketing and psychology (Balcombe, Fraser, and McSorley, 2015), but is rather new in economics, where it has found some applications in behavioral and neuroeconomics (see Lahey and Oxley (2016) for an overview). In the context of discrete choice experiments there has been a number of eye-tracking studies (see Orquin and Mueller Loose (2013) for an overview of the use of eye-tracking in decision-making studies). So far, Balcombe, Fraser, and McSorley (2015), Spinks and Mortimer (2016), and Van Loo et al. (2014) are the only studies that use eye- tracking technology to investigate ANA. If eye-tracking is used in the context of ANA, attribute attendance needs to be clearly defined as several measures are possible. Eye movements consist of fixations and saccades. Fixations are still mo- ments of focus that last for about 200 to 500 milliseconds (ms). Saccades describe quick movements of 20 to 40ms where the gaze shifts from one area to another area of interest (Balcombe, Fraser, and McSorley, 2015). Contrary to stated and inferred ANA, Balcombe, Fraser, and McSorley (2015) label ANA information obtained from eye-tracking as visual ANA. According to their definition, visual ANA occurs if an individual has less than two eye fixations on an attribute in the majority of the choice tasks. The results of their study show that stated and visual ANA, the two approaches they compare, are only weakly correlated given that respondents state that they are attending to specific attributes, but appear to pay very limited visual attention to them. They find that both stated and visual ANA improve model performance compared with the choice models that do not account for ANA, and that the model based on stated ANA performs better than the model which uses visual ANA information. Although eye-tracking provides data that are conceptually linked closely to perceptual representation (Orquin and Mueller Loose, 2013), there are some obstacles for its use in choice experiments. First, eye-tracking studies have to be conducted in a laboratory setting. Secondly, the need to closely monitor respondents and frequently recalibrate the eye-tracking device precludes tracking several respondents simultaneously (Schulte-Mecklenbeck, Murphy, and Hutzler, 2011). Due to the complexity and time requirements, eye-tracking studies tend to result in high unit costs per participant. More often than not these high study costs are compensated by reducing the number of study participants. Balcombe, Fraser, and McSorley (2015), for example, conducted their choice experiment 76 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments with only 40 respondents. This issue can be further aggravated by the loss of sample representativeness because of the need to exclude some groups of the general population in eye-tracking studies. For example, participants can typically not be much older than 65 since drooping eyelids may be an issue, and it is difficult to track respondents wearing glasses or suffering from eye diseases such as cataracts (Waechter, Sütterlin, and Siegrist, 2015). Finally, respondents are inevitably aware that their eye movements are being tracked which could bias the results. An alternative visual technique to eye-tracking is mouse-tracking. Mouse- tracking avoids most of the issues mentioned above. In particular, it can be eas- ily used in online surveys thereby facilitating comparisons with large parts of the standard choice experiment literature. Mouse-tracking is a method that uses software to monitor information acquisition processes on websites. A number of applications are available, but most of them function by covering (part of) the information provided to respondents. This information is only uncovered once the mouse cursor hovers over the area where the information is hidden, and only while the mouse cursor hovers over it. The tool usually records frequency and duration of uncovering a specific area with information hidden behind it. The simpler setup of mouse-tracking compared with eye-tracking has many advantages, but also some disadvantages. A main issue is that the recorded mouse movements only serve as a proxy for eye movements. There might be situations where respondents look away while uncovering a specific piece of information with their mouse thereby wrongly suggesting attribute attendance. Mouse behavior is, moreover, entirely under a participant’s cognitive control, whereas unconscious movements can only be recorded using eye-tracking (Reisen, Hof- frage, and Mast, 2008). However, a number of studies show that mouse-tracking provides similar results to eye-tracking, and that the differences between the methods are not substantial for various applications (Lohse and Johnson, 1996; Reisen, Hoffrage, and Mast, 2008; Schulte-Mecklenbeck, Murphy, and Hutzler, 2011). For example, Schulte-Mecklenbeck, Murphy, and Hutzler (2011) compare the mouse-tracking software Flashlight with eye-tracking. The results show a similar outcome irrespective of the tracking method used. However, information acquisition takes longer in the mouse-tracking treatment. A reason for this may be that the eye-tracking study was conducted in a laboratory setting involving students, while the mouse-tracking was conducted online with a more representative sample of the population (Schulte-Mecklenbeck, Murphy, and Hut- zler, 2011). Nevertheless, although mouse-tracking does not measure exactly the 4.3. Case-study description and experimental design 77 same thing as eye-tracking, the data obtained by mouse-tracking seem to produce results that suffice for answering many economic research questions.

4.3 Case-study description and experimental design

The conducted choice experiment elicited public preferences for a hypothetical expansion of hydropower production in Switzerland. The choice experiment design included four attributes describing two hypothetical expansion alternatives and a status quo alternative. The first attribute defines the type of expansion, i.e. the construction of new hydropower plants or the extension of existing facilities. The next two are risk attributes. One of them describes an increase in the risk of a dam breach. The other risk attribute is framed as a decrease in the risk of a nuclear accident as a result of expanding hydropower and consequently shutting down nuclear power production. Both risk attributes are defined in terms of the risk of death in an average lifetime. Risk levels are carefully communicated to respondents by means of risk ladders. The baseline risks (risks associated with the status quo alternative) for both events are calculated based on existing studies (Burgherr and Hirschberg, 2008, 2014; Hirschberg et al., 2016). These studies estimate fatalities per Gigawatt-electric-year(GWeyr) of different electricity production technologies for different regions. Using estimates for hydropower and nuclear power capacities in Switzerland in terms of GWeyr, and normalizing for population size and life expectancy, we transformed the estimates from this literature to the average lifetime risks of death. Finally, a price attribute is defined in terms of an increase in a household’s annual electricity bill. A list of the attributes and their corresponding levels is shown in Table 4.1. Public understanding of the survey and, specifically, the relevance of the choice attributes and their levels, were tested and modified based on 20 in-depth face-to-face interviews and two online pretest rounds. These online pretests were carried out among representative samples for the German- speaking Swiss population consisting of 220 and 350 participants in the first and second round, respectively. The respondents were recruited by Intervista AG, a market research company with a panel of 50,000 registered individuals throughout Switzerland. The pretests also served to test the efficiency of the choice experiment design and to derive prior estimates for the attribute coefficients. A D-efficient experimental design was generated in Ngene 1.1.2. Since the priors obtained in the pretests are not precise due to modifications made in the design of the main survey after the 78 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments

TABLE 4.1: Attributes and attribute levels in the choice experiment

Increase in household’s 100, 200, 300, 400, 500, 600 No change in the annual electricity bill annual electricity bill (CHF) pretests, we used Bayesian priors for generating the main survey design (Choice- Metrics, 2012). The Bayesian priors are specified to follow a normal distribution within the limits provided by the priors obtained from the two pretests. In total, two blocks consisting of seven choice tasks each were generated and randomly assigned among respondents. Dominant alternatives were excluded in the design generation procedure. Each choice task included two hypothetical policy alternatives and a status quo alternative. The main survey was conducted in June 2016 among a representative sample of the German- and French-speaking Swiss population, covering roughly 95% of the country’s population. A total of 247 respondents completed the survey, giving a response rate of 16.3%. The average response time for the entire survey was 24 minutes. Five of the 247 respondents were identified as fast-clickers and excluded from the analysis. These respondents took less than 10 seconds to answer every choice task page. Excluding respondents from the analysis based on the duration of answering choice questions is disputed in the literature, since it is not trivial to define what a "fast" answer is (Campbell, Mørkbak, and Olsen, 2017). In our choice experiment, where the mouse-tracking setup requires some effort to acquire the relevant information needed to take a decision, we consider 4.4. Elicitation of stated and visual ANA 79

10 seconds a reasonable threshold. Ten protest responses were furthermore iden- tiﬁed and excluded from the analysis following the recommendation in Johnston et al. (2017), yielding a sample of 232 respondents for further analysis.

4.4 Elicitation of stated and visual ANA

The main goal of this study is to compare the model performance of models that take visual, stated and inferred ANA into account. All models are based on the same sample of respondents and the same experimental design. Inferred ANA is based on the choice experiment data and derived using econometric modeling approaches (see Section 4.5), whereas stated and visual ANA are elicited in the survey. Stated ANA is elicited by asking the survey respondents the following question after the whole choice experiment: "If you think back about your choices, to what extent did you pay attention to the following characteristics?" The question was followed by a slider scale for each attribute, ranging from 0 (never considered the attribute) to 10 (always considered the attribute). This resulted in an ordinal measure of attendance which was subsequently transformed to a binary variable of stated ANA (see Section 4.6.2). Visual attendance was obtained through mouse-tracking. To track mouse movements we used the software MouselabWEB (Willemsen and Johnson, 2010). Except the attribute description, each row displaying the attribute levels across the three alternatives was thereby covered by grey boxes and only uncovered once the mouse hovered over it (see Figure 4.1 for an example of a choice task with the price attribute uncovered2). The whole attribute row was uncovered if the mouse hovered on any box in a row since we were more interested in the visual information acquisition process related to the attributes than the alternatives. The information on the attribute levels was covered again once the mouse cursor left the grey box covering the row. It was not required to click the mouse for uncovering boxes. For each uncovering event, the software records the exact time of opening a row, the duration of uncovering it (in milliseconds), and the time of closing. For our analysis we use the information about the total number of times a row was uncovered (the frequency of uncovering), as well as the total duration of uncovering an attribute row per choice task. Since there is a strong tendency for respondents to start uncovering boxes in the top-left corner, it is

2Appendix 4.A shows an example of a choice task with all attributes uncovered. Note that the survey respondents could only uncover one attribute row at the time. 80 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments common practice to counterbalance this effect by randomizing the presentation of the attributes and alternatives (Willemsen and Johnson, 2008). In this study we focused on the information acquisition process concerning attributes, and therefore only the order in which attributes were shown to respondents across choice tasks was randomized.

A) Expansion B) Expansion C) No expansion

Type of hydropower expansion New construction Extension No expansion

Risk of death from a

dam breach +40% risk +20% risk Current risk (1 in 650’000 people) (1 in 750’000 people) (1 in 900’000 people)

Risk of death from a

nuclear accident -30% risk -60% risk Current risk (1 in 4M people) (1 in 7M people) (1 in 3M people)

Increase in your household‘s yearly +200 CHF/year +300 CHF/year +0 CHF/year electricity bill

FIGURE 4.1: Example choice task as shown to respondents in the mouse-tracking setup with the price attribute uncovered

Eye-tracking studies need to define a cut-off level in order to distinguish fixations from saccades, i.e. still eye movements that allow information acquisition from quick movements of gaze shifting where no information lookup takes place (Meissner, Scholz, and Decker, 2010). Movements of the eyes do not perfectly correlate with movements of the mouse as the eye is expected to move more. Mouse movements that do not serve the acquisition of information, but happen because respondents move the mouse to an area with information content of interest and cross and uncover several boxes on the way for a very short time, represent spurious information acquisitions since no actual content is ac- quired. These movements should be excluded from the analysis. Although there exists no consensus on the cut-off point that can be applied to these situations, 4.5. Econometric models 81 acquisitions of less than 200ms can hardly be read by participants (Willemsen and Johnson, 2008). We hence exclude all uncovering events shorter than 200ms from the analysis presented here, which leaves us with roughly 25% of the initial number of events. As in the case of stated ANA, we transform the information on the frequency and the duration of uncovering attributes to a binary definition of visual ANA (see Section 4.6.3).

4.5 Econometric models

This section describes the econometric models used in the analysis. In general, the choice models that use information on ANA are the same, irrespective of whether the ANA data originate from stated responses or visual observations. The section first explains the models used for the analysis of stated and visual ANA information, followed by a description of models estimated for the analysis of inferred ANA information. We incorporate stated and visual ANA information into choice models by restricting attribute parameter estimates for respondents who do not attend to an attribute to zero for the following reasons3. First, it is the most common model specification found in the stated ANA literature, and therefore provides the most straightforward comparison with this strand of research. Secondly, this choice model is also applied in the sparse literature which investigates visual ANA by eye-tracking, allowing again for a direct comparison. Finally, this econometric approach resulted in the best model fit. Alternative ANA models described in Section 4.2, such as estimating separate coefficients for attenders and non- attenders (e.g. Alemu et al., 2013; Carlsson, Kataria, and Lampi, 2010), were also estimated, but performed worse in terms of goodness of fit, possibly due to the higher number of regressors required while using the same number of observations. The standard specifications of multinomial logit (MNL) and mixed logit (MXL) models are used, imposing the parameter restrictions for respondents who do not attend to an attribute. These choice models can be derived from McFad- den’s random utility framework (McFadden, 1974), which assumes an informational asymmetry between utility-maximizing individuals and observing ana- lysts. That is, each individual is aware of all her utility parameters, but only part of these are observable by the analyst. The standard MNL model assumes

3Note that the attribute coefficients and not the attribute levels are constrained to zero (Hensher, Rose, and Greene, 2012). 82 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments that the unobserved factors are uncorrelated across alternatives as well as choice tasks in a choice experiment. This assumption could potentially be erroneous, as it is possible that unobservable factors related to one alternative might also influence other alternatives (Train, 2009). The same may hold for the unobserved factors over a sequence of choice tasks. In addition, although the MNL model can capture taste heterogeneity between respondents across observable characteristics through the specification of explanatory variables, it does not allow for random taste variation. For these reasons, we also estimate MXL models for the analysis of stated and visual ANA. In contrast to the MNL model, the MXL model allows for random taste heterogeneity across individuals, as well as correlation between unobserved factors over alternatives and choice tasks. Formally, it is a weighted average of the standard MNL probability representation P that an individual n chooses alternative i over alternative j in choice situation t, with weights given by the density function f() (Train, 2009):

x e 0n nit Pnit()= f()d, (4.1) nxnjt j e 0 Z ⇣ ⌘ P where denotes a vector of parameters and x a set of observed variables. The density f() can follow any distribution specified by the analyst. Some authors suggest using the uniform distribution for binary attributes (e.g. Hensher and Greene, 2002), while other studies use distributions that restrict the sign of the coefficients for the price variable, e.g. by assuming a log-normal or triangular distribution. We specify the distribution of all parameters in this study as normal. Normal distribution of parameters not only ensures the best model fit, but it is also a "natural" choice as it can be motivated by the central limit theorem (Greene, 2012). Normally distributed attribute parameters are not constrained to positive or negative values and, hence, may give theoretically unexpected signs. Here we allow the distribution of the parameters to be driven by the data, given that we are specifically interested in how individual respondents attend to and weigh the individual choice attributes and not the estimation of theoretically cor- rect welfare measures. The actual elements of in equation 4.1 that are part of the log-likelihood function are restricted to zero for attributes which are ignored by a specific respondent based on either stated or visual data. Therefore, for each individual n 4.5. Econometric models 83 and attribute k, the corresponding element can be described as:

nk, if attribute k is attended to by individual n; nk = (4.2) ( 0, if attribute k is ignored by individual n.

On the basis of equations 4.1 and 4.2, the MXL model is estimated using maximum likelihood estimation on the appropriate log-likelihood function. Inferred ANA is analyzed using ECLC models in which latent classes correspond to speciﬁc ANA behavior. Although some authors advocate not restricting coefﬁcients across classes to equality (e.g. Hensher, Rose, and Greene, 2015), the advantage of constraining the parameters to be equal across classes is that other sources of heterogeneity between the classes other than ANA are excluded. In its essence, ECLC and the more general LC models are MXL models with a discrete instead of a continuous weighting function f(). Following (Train,

2009), we can assume, for example, that takes F possible values (1,...,F ). In the general case, there are therefore F classes in the population. The probability of an individual belonging to class f can be deﬁned as qf , resulting in the following model speciﬁcation:

F x e 0f nit Pnit = qf . (4.3) f xnjt j e 0 fX=1 ⇣ ⌘ P The LC models infer class membership probabilities from the observed choices made by respondents. Since the classes are defined in terms of different ANA behavior patterns, the inferred class membership probabilities represent the probability of non-attendance to different (combinations of) attributes. Total ANA to an attribute is then obtained by summing the membership probabilities across all classes in which this attribute is assumed to be not attended to. A key issue in estimating LC models is finding the appropriate number of classes. In this study we followed the "test down" procedure, as suggested by Hensher, Rose, and Greene (2015). We start with the "2k" model developed by (Hole, 2011), where 2k indicates the maximum number of combinations of ANA behavior, and k is the number of attributes. This model is estimated based on the maximum possible number of classes defined by all possible combinations of ANA patterns while at the same time ensuring model convergence. Subsequently, we gradually eliminate classes with low membership probabilities, and search for the best model based on the Bayesian and Akaike Information Criteria (BIC and AIC, respectively). All models are estimated in NLOGIT 5. 84 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments

4.6 Results

4.6.1 Descriptive statistics

The sociodemographic characteristics of the sample and of the target population are described in Table 4.2. The comparison of the sample characteristics with the German- and French-speaking Swiss population shows that the two are congru- ent, albeit there are some differences. In particular household income is slightly lower in our sample than the population average.

TABLE 4.2: Sociodemographic characteristics of the study sample and target population

Sample statistics Population statisticsa Share female (%) 49.4 50.5 Average age (years) 47.3 41.5 Share French-speaking origin (%) 24.7 23.0 Average household size 2.4 2.3 Share univ. degree (Bachelor’s, Master’s, PhD) (%) 28.3 27.1b Average annual gross household income (CHF) 97,442c 120,624 Share unemployed (%) 2.4 4.3 Share retired (%) 20.1 17.9d Share student (%) 9.3 5.7b N 247 7,887,303 Notes: aSFSO (Swiss Federal Statistical Ofﬁce) (2016). bOf population aged 25 to 64. cThis is an approximate estimation as we only asked the respondents to indicate in which income group their household falls. Furthermore, 6.5% of the respondents did not answer this question. dPeople aged 65 or older.

Figure 4.2 shows the summary statistics on stated ANA for the four choice attributes. As we asked respondents to state ANA after completion of all the choice tasks, it presents the average self-reported ANA over all choice tasks and respondents. Figure 4.2 shows that attendance to the choice attribute describing the type of hydropower expansion is highest, followed by attendance to the price attribute and the risk of a nuclear accident. The risk of a dam breach is the least- attended attribute. Table 4.3 summarizes the data on the two key indicators for visual ANA obtained from mouse-tracking across the choice task sequence: the frequency and

4Error bars reﬂect one standard deviation. 4.6. Results 85

0 Type of hydropower Risk of death from a Risk of death from a Increase in expansion dam breach nuclear accident household's annual electricity bill

FIGURE 4.2: Descriptive statistics for mean stated ANA, from 0 (never considered the attribute) to 10 (always considered the attribute)4 the duration of uncovering a choice attribute. It reports the frequency and duration that an attribute is uncovered on average across choice tasks and respondents. The mean values for frequency and duration show the highest attendance to the type of hydropower expansion, which is in line with the stated ANA results. The shortest time is dedicated to the price attribute, and the lowest frequency of uncovering the attribute is assigned to the lifetime risk of death from a nuclear accident. However, the differences in frequencies and durations between the price and the nuclear risk attributes are negligible. Note that ordering effects cannot be driving the results, as attributes were shown in a randomized order across respondents and choice tasks. The average frequency and duration of attribute uncovering for each choice task are shown in Appendix 4.B. They reveal that both indicators of visual attribute attendance decrease over the choice task sequence for all attributes: the average frequency of uncovering an attribute in the first choice task is roughly 4.35 with an average lookup duration of 11,300ms. These values are, respectively, 79% and 58% lower in the last choice task. Interestingly, the frequencies and durations of attribute-uncovering decrease most strongly for all attributes during 86 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments the first two choice tasks, and stabilize after the third choice task. The Kruskal- Wallis test procedure was applied to test the equality of attribute-specific frequencies and durations of attribute-uncovering across choice tasks. The test rejects the hypothesis of equality of frequencies and durations of attribute-uncovering across all choice tasks.

TABLE 4.3: Visual ANA statistics

Avg. lookup frequency Avg. lookup dur. in ms Attribute Mean Std. Min. Max. Mean Std. Min. Max. Dev. Dev. Type of hydropower 2.72 2.68 0 25 5250 8051 0 108475 expansion Lifetime risk of death 2.53 2.31 0 26 4414 6057 0 98876 from a dam breach Lifetime risk of death 2.14 2.06 0 17 3890 5752 0 64262 from a nuclear accident Increase in household’s 2.18 1.97 0 20 3727 5568 0 115102 yearly electricity bill

4.6.2 Stated ANA models

Stated ANA information is included in the choice models by restricting the attribute coefficients of respondents stating ANA to zero. The stated ANA information obtained from the survey is an ordinal variable which was converted to a binary one. Since there is no theoretical basis on which a cutoff point can be defined, we determine the threshold separating non-attenders from attenders using model performance indicators. For the MXL models, which proved to perform better than the MNL models, we calculate the AIC and BIC values for every possible cutoff point, as depicted in Appendix 4.C. Appendix 4.C illustrates that both the AIC and the BIC are lowest for the model that defines stated ANA as lower than 6 (on a scale from 0 to 10). This is therefore selected as the threshold value, meaning that all respondents who stated a degree of attribute attendance lower than 6 are treated as non-attenders for the specific attribute. The estimated choice models that use the cutoff point explained above are reported in Table 4.4. Since MXL models proved to perform better than MNL models in this study, we only show the results of these specifications for all the models estimated5. Model 1 assumes FAA and serves as baseline case. Model 2 takes stated ANA into account. The models were estimated using 2,500 Halton

5MNL model results are available from the authors on request. 4.6. Results 87 draws. All choice attributes except the price are specified as binary variables. The coefficient for the first attribute (type of expansion) describes the effect of constructing new hydropower plants compared with the extension of existing plants. The attribute reflecting the lifetime risk of death from a dam breach is also included as a binary variable indicating an increase in hydropower risk by 40% compared with an increase of 20%. The attribute describing the lifetime risk of death from a nuclear accident is treated analogously, with its coefficient describing the marginal effect of a risk reduction by 60% compared with a reduction of 30%. The ASC describes the effect of the status quo alternative compared with the hypothetical hydropower expansion alternatives6. The model results show that all coefficients associated with the choice attributes are highly significant and have the expected signs. Hydropower expansion scenarios are, on average, preferred to the status quo situation, as reflected by the negative coefficient for the ASC. However, the construction of new hydropower plants decreases the utility of respondents compared with extending existing plants, possibly because new constructions have stronger environmental impacts than extensions. The signs of both risk attributes are as expected as well: an increase in hydropower risk by 40% reduces the likelihood of choosing a hypothetical hydropower expansion alternative compared with an increase of 20%, while a decrease in nuclear risk by 60% increases the likelihood of choosing a hypothetical alternative compared with a decrease of 30%. An increase in the household electricity bill affects respondents’ utility negatively. The standard deviations of the random parameters are significant, with the exception of the hydropower risk attribute, which suggests heterogeneity in respondents’ preferences for all but the hydropower risk attribute. Comparing the model that does not consider ANA (Model 1) with the model accounting for stated ANA (Model 2) reveals that there is an improvement in model fit when stated ANA information is incorporated into the choice models. This result confirms findings in the literature, which shows that including stated ANA information into choice models improves model performance (e.g. Weller et al., 2014). Neither the signs of the coefficients nor the coefficients’ significance levels differ between the two models. 6Although the type of hydropower expansion and the hydropower and nuclear risk attributes have three levels, only one binary variable for each attribute was included in the analysis. The reason for this is that their status quo level was only part of the status quo alternative. The effect of a change from the status quo to another level is hence unidentified and perfectly confounded with the ASC. 88 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments

TABLE 4.4: Full attribute-attendance and stated ANA MXL models

Model 1: Model 2: FAA Stated ANA Variables Coeff. s.e. Coeff. s.e.

Mean estimates ASC -4.592*** 0.445 -4.633*** 0.456 Construction of new hydropower plants -0.896*** 0.138 -1.266*** 0.195 Increase in lifetime risk of death from a dam -0.186** 0.083 -0.609*** 0.158 breach by 40% Decrease in lifetime risk of death from a 0.527*** 0.098 0.917*** 0.154 nuclear accident by 60% Increase in household’s yearly electricity bill -0.004*** 0.001 -0.006*** 0.001 Standard deviations of random parameters ASC 3.692*** 0.411 3.752*** 0.404 Construction of new hydropower plants 1.533*** 0.152 1.916*** 0.208 Increase in lifetime risk of death from a dam 0.038 0.444 0.178 0.648 breach by 40% Decrease in lifetime risk of death from a 0.778*** 0.133 1.051*** 0.185 nuclear accident by 60% Increase in household’s yearly electricity bill 0.004*** 0.001 0.005*** 0.001 Model characteristics Number of observations 1624 1624 Log-likelihood (restricted) -1673.255 -1673.255 Log-likelihood (unrestricted) -1240.769 -1172.048 AIC/N 1.540 1.456 BIC/N 1.574 1.489 McFadden’s pseudo R2 0.305 0.343 Note: ***p<0.01, **p<0.05, *p<0.1. 4.6. Results 89

4.6.3 Visual ANA models

Defining an appropriate threshold that clearly separates respondents who attend to the attributes from those who do not is even more crucial in the case of visual ANA than in the case of stated ANA, since more definitions of ANA behavior are conceivable using the information on visual attendance. Visual ANA can be defined in terms of either frequency or duration of attribute attendance, and it can be defined in either absolute or relative terms, i.e. whether an ANA threshold is defined in absolute terms or relative to a respondents’ attendance to other attributes. Furthermore, there is the possibility to distinguish between choice-task specific or serial ANA. The main criterion for selecting the definition of visual ANA in this study is its comparability to the definition of stated ANA. This means that visual ANA must also be serial. Furthermore, we initially follow the definition of visual ANA proposed by Balcombe, Fraser, and McSorley (2015) and Van Loo et al. (2014) in the context of eye-tracking. They define visual ANA to a particular choice attribute as "ignorance of this attribute in a majority of the choice tasks". Ignorance is thereby defined as less than two fixations on the specific attribute. It is therefore an absolute definition of ANA. In our case, non-attendance to an attribute is defined as the "frequency of uncovering an attribute less than two times in a majority (four or more out of seven) of the choice tasks". The coefficients for the non-attended attributes are restricted to zero in an analogous manner as for the stated ANA models. The resulting choice model incorporating visual ANA information as defined above is presented as Model 3 in Table 4.5. The coefficients in Model 3 are similar in terms of signs, magnitudes, and significance levels to the FAA and stated ANA models (Models 1 and 2) presented earlier. The performance of the visual ANA model based on lookup frequency indicates a deterioration compared with both the FAA model (Model 1) and the model incorporating stated ANA information (Model 2). In addition to the visual ANA definition used in Balcombe, Fraser, and Mc- Sorley (2015) and Van Loo et al. (2014), both of which use eye-tracking, we present a new definition of visual ANA based on mouse-tracking data. After ex- perimenting with a large number of models using various definitions and thresholds of visual ANA, we present the best-performing model. This model defines ANA in terms of the duration of attribute-information uncovering instead of frequency. Model 4 assumes that ANA occurs if the total duration of attending to an attribute across the entire sequence of choice tasks is shorter than 10,000ms. 90 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments

TABLE 4.5: Visual ANA MXL models based on different thresholds of lookup frequency and duration

Model 3: Model 4: <2 ﬁxations in <10,000 ms across majority of choice tasks all choice tasks Variables Coeff. s.e. Coeff. s.e.

Mean estimates ASC -4.378*** 0.499 -4.686*** 0.467 Construction of new hydropower -1.191*** 0.164 -0.921*** 0.146 plants Increase in lifetime risk of death -0.208** 0.088 -0.230*** 0.087 from a dam breach by 40% Decrease in lifetime risk of death 0.521*** 0.127 0.571*** 0.102 from a nuclear accident by 60% Increase in household’s yearly -0.004*** 0.001 -0.004*** 0.001 electricity bill Standard deviations of random parameters ASC 4.176*** 0.443 3.922*** 0.412 Construction of new hydropower 1.483*** 0.175 1.595*** 0.159 plants Increase in lifetime risk of death 0.035 0.423 0.063 0.377 from a dam breach by 40% Decrease in lifetime risk of death 0.816*** 0.174 0.778*** 0.142 from a nuclear accident by 60% Increase in household’s yearly 0.005*** 0.001 0.005*** 0.001 electricity bill Model characteristics Number of observations 1624 1624 Log-likelihood (restricted) -1673.255 -1673.255 Log-likelihood (unrestricted) -1271.914 -1235.730 AIC/N 1.579 1.534 BIC/N 1.612 1.567 McFadden’s pseudo R2 0.287 0.307 Note: ***p<0.01, **p<0.05, *p<0.1. 4.6. Results 91

Model 4 that uses lookup duration to deﬁne ANA performs better than the model deﬁning visual ANA based on the frequency of information acquisition. Nevertheless, the improvements relative to the baseline FAA model are minor. Comparing the visual ANA model (Model 4) with the model that relies on stated ANA information (Model 2), it is evident that the stated ANA model performs better.

4.6.4 Inferred ANA models

LC models can be used to infer the patterns of ANA behavior from the data structure. In contrast to the standard LC model, the latent classes in an ANA context represent specific behavioral strategies with respect to ANA. The estimated ECLC model is shown in Table 4.6. Model 5 is the best-performing ECLC model, and is based on the test-down procedure (Hensher, Rose, and Greene, 2015). This choice model allows for combinations of non-attendance to several attributes7. Model 5 does not allow for taste heterogeneity within and between classes. We also ran ECLC models allowing for random taste parameters. However, these models only converged for a subset of ANA classes. They also resulted in only very few significant class membership probabilities. Finally, we also estimated models without constraining the coefficients to be equal across classes. How- ever, in most cases they show what Hensher, Rose, and Greene (2015) refer to as "signature features of overfitting": namely, large standard errors, very small class membership probabilities, and large differences in the coefficients across classes. The estimated attribute coefficients are only shown for one class in Table 4.6, as they are the same for the attended attributes across all classes. The estimated model parameters have the same signs, and comparable significance levels to most of the stated and visual ANA models presented in Tables 4.4 and 4.5. With respect to class membership probabilities, there are two remarkable results. First, the probability of belonging to the FAA class is statistically insignificant. Secondly, the probability assigned to class 7, which is characterized by non-attendance to all attributes, is relatively high. Comparing the inferred ANA model with the stated and visual ANA models, it follows that the ECLC model performs less than the models using stated and visual ANA information. The shares of respondents exhibiting ANA behavior based on stated, inferred, and mouse-tracking data are compared in Table 4.7. They differ between

7An alternative model, which restricts ANA behavior to non-attendance to a single attribute and FAA behavior to all four choice attributes, is presented in Appendix 4.D. This model does not include an option of non-attendance to more than one attribute. 92 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments

TABLE 4.6: Inferred ECLC ANA model

Model 5: ECLC model Variables Coeff. s.e.

Mean estimates ASC -2.148*** 0.123 Construction of new hydropower plants (a) -3.750*** 0.427 Increase of lifetime risk of death from a dam breach by 40% (b) -1.392*** 0.316 Decrease of lifetime risk of death from a nuclear accident by 60% (c) 2.091*** 0.324 Increase in household’s yearly electricity bill (d) -0.008*** 0.000 Class membership probabilities Class 1: FAA 0.000 0.020 Class 2: ANA to (a) 0.048** 0.024 Class 3: ANA to (c) 0.116*** 0.032 Class 4: ANA to (a), (d) 0.117*** 0.032 Class 5: ANA to (b), (d) 0.238*** 0.033 Class 6: ANA to (a), (b), (c) 0.113*** 0.030 Class 7: ANA to (a), (b), (c), (d) 0.368*** 0.040 Model characteristics Number of observations 1624 Log-likelihood (restricted) -1673.255 Log-likelihood (unrestricted) -1312.405 AIC/N 1.630 BIC/N 1.666 McFadden’s pseudo R2 0.264 Note: ***p<0.01, **p<0.05, *p<0.1. 4.6. Results 93 the ANA approaches, as they are dependent on the definition of ANA or, more specifically, on the cutoff point used for stated and visual ANA definition and the number and characteristics of classes used for the inferred ANA model. There is also no observable clear pattern in relative terms, i.e. when comparing the most and least attended attributes between the different ANA definitions. Based on stated ANA and visual ANA lookup duration, the type of hydropower expansion is the most-attended attribute, whereas in the case of inferred ANA the lifetime risk of death from a nuclear accident is the most-attended attribute. Ac- cording to the stated and visual ANA approaches the respondents display the highest propensity of non-attendance to the risk attributes: hydropower risk in the case of stated ANA, and nuclear risk in the case of visual ANA. Again, the result for inferred ANA is different, and shows that the price attribute is the least attended attribute in the best-fit model. The substantial differences in the shares of respondents exhibiting ANA behavior between the different ANA approaches are in line with the results found in the literature comparing stated and inferred ANA (e.g. Campbell, Hensher, and Scarpa, 2011; Kragt, 2013; Weller et al., 2014).

TABLE 4.7: Shares of respondents displaying stated, visual, and inferred ANA

Type of Lifetime Lifetime Increase in hydropower risk of death risk of death household’s expansion from a dam from a annual breach nuclear electricity accident bill ANA deﬁnition Share of respondents

Stated ANA Stated ANA <6 29.3% 72.4% 49.1% 40.9% Visual ANA <2 ﬁxations in a majority of 29.7% 27.6% 44.8% 37.5% choice tasks <10,000ms uncovering 3.4% 7.3% 7.8% 7.3% duration across all choice tasks Inferred ANA ECLC model 64.6% 71.9% 59.7% 72.3% 94 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments

4.7 Discussion and conclusions

A comparison of the model performance indicators across the estimated models enables us to draw several conclusions. First, taking stated or visual ANA into account improves the model performance compared with the conventional choice models assuming FAA. Second, comparing the models based on the stated and inferred ANA approaches shows that the stated ANA model outperforms the inferred model. Third, although the visual ANA model that uses a duration- based ANA definition performs slightly better than the baseline FAA model and better than the inferred ANA model, it performs worse than the stated ANA model. Therefore, we have to reject our central hypothesis that including visual ANA information results in improved choice models compared with existing ANA models in the literature. Although this is somewhat contrary to expectations, as revealed ANA behavior was expected to produce more reliable information than stated ANA behavior, and hence increase the explanatory and predictive power of the choice models, this finding does not contradict the limited evidence in the existing literature using eye-tracking to study ANA. Balcombe, Fraser, and McSorley (2015) find that a model based on stated ANA performs better than a model based on visual ANA, and also in their case both ANA models perform better than the FAA model which does not account for ANA. Van Loo et al. (2014) use the same definition of visual ANA as Balcombe, Fraser, and McSorley (2015) and introduce two additional visual ANA definitions, but none of them improves the baseline FAA model. Interestingly, other studies which compare eye-tracking information with stated responses outside the domain of ANA find similar results. Uggeldahl et al. (2016), for example, compare choice models taking stated choice certainty into account with models that consider choice certainty based on eye-tracking. They also find that the inclusion of stated information results in better models than the inclusion of visual information. The results of our study are hence in line with the findings in the existing literature on visual ANA and in other literature on eye-tracking. Note, however, that the studies on eye-tracking ANA have very low sample sizes of 81 and 40 participants in Balcombe, Fraser, and McSorley (2015) and Van Loo et al. (2014), respectively. To our knowledge, there is no study yet that uses mouse-tracking to investigate ANA with which we could compare our results. There may be several explanations for the results found in visual ANA studies in the literature including our study. First, it is not clear what is the most appropriate definition and measurement of visual ANA behavior, independently 4.7. Discussion and conclusions 95 of whether it is examined by eye- or mouse-tracking techniques. There are nu- merous possible definitions based on either the frequency or the duration of attribute attendance, which can furthermore be measured either in absolute or relative terms. ANA behavior can also be defined as choice-task specific or serial. We tried a number of alternative definitions, but none of them resulted in a substantial model improvement. Moreover, the results obtained from models which use different definitions of visual ANA are rather similar, suggesting a low sensitivity to varying ANA definitions. We observe that more stringent definitions of visual ANA, i.e. labeling more respondents as visual non-attenders, often resulted in lower model performance. This may be an indication that the relationship between visual and true ANA is not as straightforward as it seems. The premise that low visual attendance to an attribute, defined either in terms of low frequency or duration of attribute lookup, represents a real behavioral non-attendance to an attribute may not always hold. That is, a low frequency or short duration of uncovering attribute information does not necessarily imply that participants in a choice experiment ignore or attach low importance to the attribute. In fact, it is plausible that the relationship may in some cases even be the reverse, meaning that respondents look at an attribute briefly and only a few times because they assign a high importance to this attribute and consequently do not require a high mental effort to process the presented information associated with that attribute. Our results suggest that ANA needs to be considered in choice models, but that it may suffice to include ANA information stated by respondents during the survey. While the former finding is in line with the existing literature, the latter is contrary to what most studies report that compare stated and inferred ANA approaches. We did not detect evidence to support the lack of trust to use stated ANA information in discrete choice models often reported in the literature. A possible explanation for this may be that in this study respondents were asked to state the strength of ANA on an 11-point measurement scale. A cutoff value that separates attenders from non-attenders was carefully defined based on different model performance indicators. This procedure seems to increase the quality of stated ANA responses in contrast to the more standard approach of asking respondents whether or not they attended to an attribute. Although we find that stated (and visual) ANA approaches provide a better model fit than the inferred ANA approach, it is important to keep in mind that the econometric models differ in the way they treat preference heterogeneity, i.e. inferred from LC or stated in MXL models. This too may have influenced the comparability of 96 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments model performance. Finally, the results may have been affected by the specific mouse-tracking procedure used in this study to assess attribute attendance and the additional cognitive burden mouse-tracking might have imposed compared to a standard choice experiment setting. The same survey data were therefore also collected from another sample of 250 respondents without the mouse-tracking treatment. The results of this comparison show that there are no statistically significant differences between the two samples in terms of their sociodemographic characteristics and the estimated choice models’ preference and scale parameters (following the Swait and Louviere (1993) test procedure, the outcome of the Likelihood Ratio test statistics are 12.33 and 1.41 for the first and second Swait-Louviere test stage, respectively, with corresponding p-values of 0.34 and 0.23). This suggests the absence of procedural bias that could have influenced the results presented here, and that choice behavior in the specific mouse-tracking procedure did not significantly deviate from choice behavior under regular online choice experiment conditions. Nevertheless, more research is needed to determine how exactly ANA manifests itself in terms of visual information acquisition, and how such information can be appropriately included in choice models. Although this study applies a sensitivity analysis by estimating models using different definitions of visual ANA, a more stringent definition of visual ANA that is well supported by visual information acquisition studies would substantially benefit future research on this topic. In addition, eye-tracking studies with larger sample sizes could help to arrive at more robust conclusions on visual ANA using this method. Comparing mouse-tracking with eye-tracking technologies using split-samples could yield additional insights into the comparative performance of the two visual tracking methods for the analysis of ANA. 4.A. Example choice task 97

4.A Example choice task

A) Expansion B) Expansion C) No expansion

Type of hydropower

expansion New construction Extension No expansion

Risk of dying from a

dam breach +40% risk +20% risk Current risk (1 in 650,000 people) (1 in 750,000 people) (1 in 900,000 people)

Risk of dying from a

-30% risk -60% risk nuclear accident Current risk (1 in 4,000,000 people) (1 in 7,000,000 people) (1 in 3,000,000 people)

Increase in your household‘s yearly +200 CHF/year +300 CHF/year +0 CHF/year electricity bill

FIGURE 4.3: Example choice task (all attributes uncovered) 98 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments

4.B Lookup frequency and duration

5 Type of hydropower expansion Lifetime risk of death from a dam breach 4 Lifetime risk of death from a nuclear accident Increase in household’s yearly electricity bill 3

Lookup Lookup frequency 2

1 Task 1 Task 2 Task 3 Task 4 Task 5 Task 6 Task 7 Choice task sequence

FIGURE 4.4: Average lookup frequency over choice tasks

16000

13500 Type of hydropower expansion Lifetime risk of death from a dam breach 11000 Lifetime risk of death from a nuclear accident 8500 Increase in household’s yearly electricity bill

6000

Lookup Lookup duration(ms) 3500

1000 Task 1 Task 2 Task 3 Task 4 Task 5 Task 6 Task 7 Choice task sequence

FIGURE 4.5: Average lookup duration over choice tasks 4.C. AIC and BIC 99

4.C AIC and BIC

1.65 AIC/N BIC/N

1.6

1.55 AIC/N; BIC/N 1.5

1.45 ANA <1 ANA <2 ANA <3 ANA <4 ANA <5 ANA <6 ANA <7 ANA <8 ANA <9 ANA<10

FIGURE 4.6: Akaike and Bayesian information criteria for MXL models based on different deﬁnitions of stated ANA 100 Chapter 4. Attribute non-Attendance in Discrete Choice Experiments

4.D Baseline ECLC model

TABLE 4.8: Baseline ECLC model

Model 6: Baseline ECLC model Variables Coeff. s.e.

Mean estimates ASC -1.888*** 0.145 Construction of new hydropower plants (a) -2.591*** 0.259 Increase of lifetime risk of death from a dam breach by 40% (b) -0.089 0.072 Decrease of lifetime risk of death from a nuclear accident by 60% (c) 0.725*** 0.058 Increase in household’s yearly electricity bill (d) -0.004*** 0.000 Class membership probabilities Class 1: FAA 0.000 0.975 Class 2: ANA to (a) 0.605*** 0.057 Class 3: ANA to (b) 0.000 0.961 Class 4: ANA to (c) 0.172** 0.086 Class 5: ANA to (d) 0.224*** 0.048 Model characteristics Number of observations 1624 Log-likelihood (restricted) -1673.255 Log-likelihood (unrestricted) -1458.266 AIC/N 1.807 BIC/N 1.837 McFadden’s pseudo R2 0.183 Note: ***p<0.01, **p<0.05, *p<0.1. 101

Chapter 5

Reference Points for the Valuation of Risk Changes in Discrete Choice Experiments1

5.1 Introduction

Prospect theory and its variant, cumulative prospect theory, are the most established choice theories that accommodate empirical findings which contradict expected utility theory (Kahneman and Tversky, 1979; Tversky and Kahneman, 1991, 1992). The main elements of prospect theory are: (1) the distinction between an editing and a valuation phase in the choice process; (2) the s-shaped utility curve that determines the utility of an outcome; and (3) a weighting function that assigns weights to risky outcomes (Stommel, 2013; Tversky and Kahne- man, 1992). The s-shaped utility curve is defined relative to a reference point that is characterized during the editing phase. It is concave for gains and convex for losses, implying diminishing sensitivity to both, and it is steeper for losses than for gains, reflecting loss aversion (Kahneman and Tversky, 1979). Some argue that dependency on reference points, i.e. stimuli that other stimuli are seen in relation to (Rosch, 1975), is the most novel element of prospect theory compared with earlier theories of choice behavior (Wakker, 2010). The general acknowledgment of the importance of reference points, and the extensively investigated question of how people react to changes from an assumed reference point, stands in contrast to the lack of theory that explains the formation and determination of reference points (e.g. Koszegi and Rabin, 2006;

1This chapter has been submitted to a journal for review. 102 Chapter 5. Reference Points for the Valuation of Risk Changes

Schmidt, 2003; Stommel, 2013). Most of the literature focuses on the effect of a specific reference point on choice. The reference point is usually defined in advance and assumed to be exogenously given (Stommel, 2013). A common problem of studies on reference points is that, in general, the reference point is unknown. Even if a hypothesis on the nature of a reference point exists in a specific situation, it may be difficult to empirically test for it, since the reference point itself is usually not directly observable. Most of the studies furthermore simply assume that the reference point equals the status quo (Arkes et al., 2008; Chen and Rao, 2002; De Moraes Ramos, Daamen, and Hoogendoorn, 2013; Genesove and Mayer, 2001; Hansson and Lagerkvist, 2014; He and Zhou, 2014; Hess, Rose, and Hensher, 2008; Heyman et al., 2004; Koetse and Brouwer, 2015; Odean, 1998; Seiler and Luchtenberg, 2014; Weber and Camerer, 1998). More recent theoretical advances focus on reference points that are based on subjective expectations about outcomes (Koszegi and Rabin, 2006), although a majority of the empirical studies that investigate expectations about outcomes precede this theoretical foundation (Abeler et al., 2011; Bartol and Martin, 1998; Chapman, 2000; Craw- ford and Meng, 2011; Hack and Bieberstein, 2014; Medvec, Madey, and Gilovich, 1995; Mellers, Schwartz, and Ritov, 1999; Ordóñez, 1998; Post et al., 2008; Saqib and Chan, 2015; Winer, 1986). Other studies define reference points as, among others, aspiration levels (Bogliacino and González-Gallo, 2015; Camerer et al., 1997; Fehr and Goette, 2007; Hack, Bieberstein, and Kraiczy, 2015; Heath, Hud- dart, and Lang, 1999; Heath, Larrick, and Wu, 1999; Lant and Mezias, 1992; Park, 2007; Sullivan and Kida, 1995); social comparisons (Fafchamps, Kebede, and Zizzo, 2015; Fiegenbaum, 1990); historical peaks (Annaert et al., 2008; Baker, Pan, and Wurgler, 2010; Gneezy, 2005; Phillips and Pohl, 2014); norms and ideals (Gómez-Mejía et al., 2007); or self-esteem (Gau and Viswanathan, 2008). This study explores the existence of reference points other than the status quo in the context of a discrete choice experiment (DCE) involving the valuation of risk changes. Reference points in DCEs have frequently been studied in terms of the specifications of attribute levels and choice alternatives. Some studies, for example, have focused on attribute levels that are pivoted around an actual experience of respondents (e.g. Hess, Rose, and Hensher, 2008; Rose et al., 2008). Another stream of literature has examined the effect of a monetary attribute that is framed as either a cost or a gain (e.g. De Borger and Fosgerau, 2008; Koetse and Brouwer, 2015; Lanz et al., 2010; Stathopoulos and Hess, 2012; Viscusi and Huber, 2012). In contrast to the bulk of the literature, our study explores the importance of reference points that are induced independently before the actual 5.1. Introduction 103 choice task. More specifically, we display two different risk ladders to two samples of respondents before asking them to value changes in risks in a DCE. We postulate that design features of these risk ladders serve as reference points. Risk ladders depict the risk in question together with other risks, covering different magnitudes of risk probabilities. These other risks serve for comparison and facilitate the assessment of the risk in question. The use of risk ladders, and visual aids more generally, for the communication of risk probabilities to survey respondents is underresearched in the context of DCEs (Logar and Brouwer, 2017). It has been shown that graphical displays of actual risk levels enhance risk understanding, especially for risks with low probabilities (e.g. Bier, 2001; Viscusi, 1998). Various devices for the visual representation of risk have been tested in the stated preference literature (e.g. Corso, Hammitt, and Graham, 2001; Dekker et al., 2011; Lipkus and Hollands, 1999; Loomis and DuVair, 1993), of which the use of a risk ladder proved to be one of the most promising in terms of providing information on relative risks (e.g. Botzen and Bergh, 2012; Loomis and DuVair, 1993). Only a few studies examine the effect of different characteristics of risk ladders. Examples include Keller, Siegrist, and Visschers (2009), who study the effect of risk ladders that contain comparative risks compared with ladders that do not, and Sandman, Weinstein, and Miller (1994), who examine the effect of the relative location of the risk in question on the ladder. Although the latter authors found this effect to be significant, they were unable to distinguish this location effect from the effect of using different comparative risk levels, since both the location of the risky events and the range of comparative risk probabilities were altered simultaneously. Our study setup allows us to isolate the effect of using different ranges of probabilities of comparative risks in risk ladders. The status quo probabilities of the risks to be valued are the same for both ladders that are shown to respondents, while the range of probabilities of the comparative risks varies. This enables us to attribute any possible differences between the two samples of respondents to the ranges of probabilities of comparative risks shown on the risk ladders, while controlling for the status quo risk level as a reference point. In this sense, the experiment tests for reference points other than the status quo. We hypothesize that other, comparative risks shown on the risk ladders serve as reference points, and hence argue for the existence of multiple reference points besides the status quo risk level of the actual risk valued. The remainder of this study is structured as follows: Section 5.2 explains the theoretical background and specifies the hypotheses. Section 5.3 describes the 104 Chapter 5. Reference Points for the Valuation of Risk Changes case study and the DCE design. Section 5.4 presents the econometric modeling framework and Section 5.5 the results. Finally, Section 5.6 closes with a discussion and conclusions.

5.2 Theoretical framework and hypotheses

Following the original text of Kahneman and Tversky (1979), which states that reference points do not necessarily have to coincide with the status quo, but can be influenced by offered prospects, expectations, comparisons and contextual factors (Carter and McBride, 2013), we posit that the comparative risks shown in a risk ladder serve as reference points, as suggested by Keller, Siegrist, and Visschers (2009) and Lipkus and Hollands (1999). Therefore, we expect that also the comparative risks influence individuals’ perception of the risk in question and hence the welfare estimates for a change in this risk. Figure 5.1 schematically illustrates how different risk ladders may provide different reference points. The risk ladders in Figure 5.1 differ with respect to the range of comparative risks they display. Risk ladder (a) in Figure 5.1 has a wide range of comparative risks and reaches a high risk endpoint (the risk shown on top of the ladder), whereas risk ladder (b) has a narrow range of comparative risks and only rises to a low risk endpoint. Both ladders start from the same risk levels depicted at the bottom of the ladders. Since we argue that comparative risks serve as reference points, we expect to find differences in public preferences and welfare estimates of the risks in question between individuals who are subjected to a risk ladder as in (a) and respondents presented with a ladder as in (b). A comparison of the two risk ladder treatments allows us to isolate the effect of comparative risks from the positioning of the risks that are to be valued in the risk ladders, which are kept constant in both risk ladders. This is in contrast to the experiment by Sandman, Weinstein, and Miller (1994), who simultaneously change the range of risk probabilities shown in the ladder and the location of the risks that were valued. Figure 5.2 illustrates the effects that different risk ladders, as shown in Figure 5.1, are expected to have on utility. Graph (a) represents the situation in which the status quo is used as a reference point, and graphs (b) and (c) show the cases of high and low comparative risk probabilities which serve as reference points. The graphs in Figure 5.2 depict our main hypothesis for the case of an increase and a reduction in risk that is to be valued. The main driver of differences between the graphs is the position of the reference point, i.e. the intersection of the 5.2. Theoretical framework and hypotheses 105

(a) Risk ladder with a high (b) Risk ladder with a low reference point reference point

High risk Low risk risk risk in in Comparative Comparative risks risks Increase Increase

S.Q. prob. of the S.Q. prob. of the valued ris ks valued ris ks Very low risk Very low risk

FIGURE 5.1: Schematic illustration of two risk ladders that differ in the range of comparative risks 106 Chapter 5. Reference Points for the Valuation of Risk Changes

same s-shaped utility curve with the x-axis (RPSQ, RPH , RPL). Although the reference points separate gains from losses in all graphs, their magnitudes, i.e. the associated risk probabilities, differ between the graphs. Graph (a) depicts the changes in utility that are expected under the traditional assumption that the status quo of the risk to be valued serves as a reference point (RPSQ). An increase in risk from RPSQ to r1 results in a deterioration compared with the status quo, i.e. the increase in risk is perceived as a loss. The opposite holds for a decrease in risk from RPSQ to r2. The latter is a positive change from the SQ reference point, and is therefore located in the gains domain. In other words, such a reference point divides not only mental gains from losses, but also the actual risk gains (decrease in risk) from the actual risk losses (increase in risk). The differences in the risk ladders as schematically shown in Figure 5.1 are not expected to affect the utility in graph (a) of Figure 5.2, since the ladders do not differ with respect to the status quo levels of the valued risks. Our alternative hypothesis is schematically depicted in graphs (b) and (c) in Figure 5.2. The comparative risks of the risk ladders are assumed to serve as reference points for the valuation of changes in the risks of importance. High and low comparative risks are hence expected to have different impacts on utility. UH in graph (b) represents the utility function that corresponds to (a risk ladder with) a high risk reference point (RPH ) (such as risk ladder (a) in Figure 5.1). A decrease in risk from SQ to r2 and an increase in risk from SQ to r1 result in a change in utility equal to the magnitude u u and u u , respectively. In contrast, U in graph (c) 3 4 4 5 L is associated with (a risk ladder with) a low risk reference point (RPL), that is, the risk has a relatively low probability of occurrence (such as risk ladder (b) in Figure 5.1). In this case, a decrease and an increase in risk result in a change of utility of u u and u u , respectively. The comparison of graphs (b) and (c) 6 7 7 8 shows that equal changes in risk (r SQ and SQ r ) are expected to result in 1 2 stronger effects on utility in the case of a low risk reference point compared with a high risk reference point, i.e. u u >u u and u u >u u . 6 7 3 4 7 8 4 5 In order to empirically assess this effect, we ﬁrst test for equality of the entire vector of estimated preference parameters between the samples using the Swait- Louviere test procedure: 1 ˆ ˆ H0 : RPH = RPL . (5.1)

More speciﬁcally, a larger change in utility is expected to be reﬂected in a higher marginal willingness-to-pay (MWTP) for a change in risk. Therefore, we subsequently also test the hypothesis, applying the Poe, Girarud, and Loomis (2005) test procedure, that the wider the range of probabilities of comparative risks on 5.2. Theoretical framework and hypotheses 107

Utility

USQ

(a) u1

Status quo Increase r1 r2 Decrease risk reference in risk RPSQ in risk

point (RPSQ)

u3 UH u4 u5

(b)

High risk Increase r2 Decrease reference point in risk RPH r1 SQ in risk

(RPH)

u6 u7 (c) u8 Low risk Increase r2 Decrease reference point in risk RPL r1 SQ in risk

(RPL)

FIGURE 5.2: Expected changes in utility associated with a status quo reference point (a), a high risk reference point (b), and a low risk reference point (c) 108 Chapter 5. Reference Points for the Valuation of Risk Changes a risk ladder, with the position of the status quo risk remaining ﬁxed, the lower the MWTP for a change in risk, ceteris paribus. The null hypothesis is speciﬁed as follows: H2 : MWTP MWTP . (5.2) 0 RPH RPL

5.3 Case-study description

5.3.1 Choice experiment

An online DCE was developed that was directly related to Swiss energy policy which aims to replace nuclear power with other energy sources, including hydropower (SFOE, 2012). The DCE pivots on the external effects of a hypothetical expansion of hydropower, which is the most established renewable source of electricity in the country. The set of choice attributes reflects direct environmental externalities associated with hydropower, and indirect external effects caused by the phasing out of nuclear power. The DCE design includes three choice alternatives: two hydropower expansion scenarios and a status quo alternative. The expansion scenarios are unlabeled. The alternatives are characterized by four attributes with the corresponding attribute levels as summarized in Table 5.1. The first attribute describes the type of hydropower expansion, i.e. whether the expansion is based on constructing new hydropower plants or extending existing facilities. This attribute is closely related to the negative environmental impacts of hydropower. It was explained to respondents that new constructions result in stronger environmental externalities than the extension of existing facilities. Two attributes describe the expected lifetime risk of dying resulting from an accident in an electricity production facility. The first is the (increase in the) risk of a dam breach that would result if hydropower electricity production were expanded. It was explained to the respondents that the risk of a dam breach would be higher if the expansion were based on hydropower plants involving only dams (40% increase), whereas a smaller risk increase would result if the expansion were based on a combination of plants with and without dams (20% increase). The second risk attribute is the (decrease in the) lifetime risk from a nuclear accident, which is an indirect externality of expanding hydropower production. Again, two levels are included in the expansion alternatives. A decrease in nuclear risk by 30% implies that the three oldest (out of a total of five) nuclear power stations in Switzerland are switched off. A decrease in nuclear risk by 60% 5.3. Case-study description 109

TABLE 5.1: Attributes and attribute levels in the DCE

Increase in household’s 100, 200, 300, 400, 500, 600 No change in the annual electricity bill annual electricity bill (CHF) means that all ﬁve Swiss nuclear power reactors are switched off. Since there are two reactors in France that are located within a distance of less than 40km to the Swiss border, the resulting lifetime risk levels from a nuclear accident do not reduce to zero. The lifetime risk levels are calculated based on risk estimates provided by Burgherr and Hirschberg (2008, 2014) and Hirschberg et al. (2016). Adjusting these estimates for the GWyr of electricity produced in Switzerland by hydropower and nuclear power, as well as for the size and life expectancy of the Swiss population, gives the estimates of the lifetime risk of dying for an average Swiss person from a hydropower or a nuclear power accident. The ﬁnal attribute is an increase in a household’s yearly electricity bill, with levels ranging from 100 Swiss Francs (CHF) to 600 CHF2. Figure 5.3 shows an example of a choice task.

5.3.2 Risk ladders

Figure 5.4 shows the two actual risk ladders that were developed and presented to the two samples of respondents in order to enhance their understanding of the current risks of a dam breach and a nuclear accident. The ladders were shown 21 CHF equaled roughly 1 United States Dollar (USD) in 2016 (OECD, 2017). The average annual electricity bill per household in Switzerland was 930 CHF in 2015 (Elcom, 2014). 110 Chapter 5. Reference Points for the Valuation of Risk Changes

A) Expansion B) Expansion C) No expansion

Type of hydropower

expansion New construction Extension No expansion

Risk of dying from a

dam breach +40% risk +20% risk Current risk (1 in 650,000 people) (1 in 750,000 people) (1 in 900,000 people)

Risk of dying from a

-30% risk -60% risk nuclear accident Current risk (1 in 4,000,000 people) (1 in 7,000,000 people) (1 in 3,000,000 people)

Increase in your household‘s yearly +200 CHF/year +300 CHF/year +0 CHF/year electricity bill

FIGURE 5.3: Choice task example 5.3. Case-study description 111 before the DCE and used in combination with textual information, as this has proven to be one of the most effective forms to communicate the risks involved (Connelly and Knuth, 1998). The textual information explained that the risks shown on the risk ladder are scientiﬁc estimates of average risks that consider both the immediate and the long-term effects. The same risk ladder was shown to respondents three times: (i) when the current risk level of a nuclear accident and a dam failure was introduced; (ii) when the change in hydropower risk resulting from a hydropower expansion was explained; and (iii) again when the change in nuclear risk due to such an expansion was described. In the last two cases, an arrow next to the current hydropower or nuclear risks indicated the direction of a risk change in case a hydropower expansion would take place, as shown in Figure 5.4.

(a) Risk ladder with a high (b) Risk ladder with a low reference point (sample 1) reference point (sample 2)

Risk of death from Risk of death from 1 in 10 people 1 in 35,000 people a 1 in 10 cancer 1 in 35,000 lightning strike

1 in 100

1 in 100,000 1 in 1000 Risk of death from 1 in 150,000 people a severe earthquake 1 in 10,000 Risk of death from 1 in 35,000 people a lightning strike

1 in 100,000

Risk of death from Risk of death from 1 in 900,000 people a dam failure 1 in 900,000 people a dam failure 1 in 1 Million 1 in 1 Million

1 in 3 Mil people Risk of death from 1 in 3 Mil people Risk of death from a nuclear accident a nuclear accident

FIGURE 5.4: Risk ladders shown to sample 1 (a) and sample 2 (b)

Since we aim to isolate the effect of the range of probabilities of the comparative risks presented on the risk ladders, half of the respondents received risk ladder (a) in Figure 5.4 (sample 1) and the other half was shown risk ladder (b) in Figure 5.4 (sample 2). Both risk ladders apply a logarithmic scale, and show 112 Chapter 5. Reference Points for the Valuation of Risk Changes the lifetime risk from a dam failure and a nuclear accident in the same location, i.e. at the bottom of the ladder. Both also compare the status quo risks of dying from a dam failure and a nuclear accident with two other commonly known risks, one of which is positioned at the top of the risk ladder (endpoint risk) and another at an intermediate level between the upper endpoint and the status quo levels of hydropower and nuclear risks (middle point risk). The main difference between the two risk ladders is the range of probabilities of these comparative risks. The endpoint risk probabilities on the risk ladders are 1 in 10 and 1 in 35,000 in sample 1 and 2, respectively, while the midpoint risk probabilities are 1 in 35,000 and 1 in 150,0003, respectively. The comparative risks on the two risk ladders differ with respect to the nature of the risks involved. In order to ensure that the risk ladders are comparable and that the ﬁndings can be traced back to the intended treatment effect, i.e. the differences in the range of probabilities of the comparative risks, we carefully selected comparative risks that are as similar as possible in terms of other factors that explain public risk perception. One such key determinant is controllability (Slovic, 1987). We control for this by asking respondents about their perceived degree of controllability of the comparative risks shown on the risk ladders. Al- though there is some evidence in the value of statistical life (VSL) literature that respondents may be willing to pay a cancer premium (e.g. Alberini and Šˇcasný, 2011, 2013; Van Houtven, Sullivan, and Dockins, 2008; Viscusi, Huber, and Bell, 2014), just as many studies have failed to ﬁnd such an effect or found mixed results (see McDonald et al. (2016) or Tsuge, Kishimoto, and Takeuchi (2005) for a review). Cancer was included as a high endpoint risk because it is commonly known. Hence, with the exception of the two different risk ladders, all other features of the DCE and the accompanying survey were identical.

5.3.3 Covariates

In addition to the stated choice data and respondents’ sociodemographic characteristics and environmental behavior, and following Hartmann et al. (2013), the survey gathered information on risk attitudes, perceptions and fears of hydropower and nuclear power accidents. Hartmann et al. (2013) link perceived threat level, coping efﬁcacy, fear arousal, and fear control associated with opposition to nuclear power and support for green energy based on the psychological frameworks of Protection Motivation Theory (PMT) (Maddux and Rogers, 1983; Rogers, 1983), the Extended Parallel Processing Model (EPPM) (Witte, 1992), and

3These risks were calculated based on data from the Swiss Federal Statistical Office (2014), the Swiss Seismological Service (2016), and a review paper on lightning risks (Ritenour et al., 2008). 5.3. Case-study description 113 theories including affective drivers (e.g. Dillard, 1994; Peters and Slovic, 1996; Slovic et al., 2007). There exists a large number of other studies focusing on the relationship between the perceived threat of nuclear power and nuclear power adoption (e.g. Greenberg and Truelove, 2011; Hartmann et al., 2013; Stoutenbor- ough, Sturgess, and Vedlitz, 2013; Tanaka, 2004; Visschers, Keller, and Siegrist, 2011; Whitfield et al., 2009) and the role of affect associated with nuclear power acceptance (e.g. Finucane et al., 2000; Peters and Slovic, 1996, 2007; Slovic et al., 2007; Visschers, Keller, and Siegrist, 2011). However, Hartmann et al. (2013) is the only study we know that considers the relationship between attitudes, perceptions and fear associated with nuclear power and the preferences for renewable sources of energy. Contrary to Hartmann et al. (2013), we assess the influence of these factors on hydropower expansion choices in a DCE. The most important variables that are used in the choice models are described in Table 5.2.

5.3.4 Design generation and data collection

A series of pretests were conducted before the main survey. First, a sample of 20 participants were asked to answer a paper-and-pencil version of the survey followed by a personal interview. In the second stage of the pretest series, a representative sample of 220 respondents from the German-speaking part of Switzer- land were asked to complete a first version of the online survey, followed by a second final online pretest round with a new sample of 350 respondents. The respondents for the pretests and the main survey were recruited by Intervista AG, a market research company with a panel of 50,000 registered individuals throughout Switzerland. Apart from testing public understanding of the survey as a whole, and the DCE in particular, the main goal of the online pretests was to derive prior estimates for the attribute coefficients. A D-efficient DCE design was generated for the final survey in Ngene version 1.1.2. Efficient designs have been shown to be superior to orthogonal designs in terms of either improving the reliability of the estimated parameters given a certain sample size or reducing the sample size needed for a given reliability of parameter estimates (e.g. Hensher, Rose, and Greene, 2015; Rose and Bliemer, 2013). Bayesian priors were used, which follow a random instead of a fixed distribution, bounded by the priors obtained from the pretests. In total, two blocks 114 Chapter 5. Reference Points for the Valuation of Risk Changes

TABLE 5.2: Explanatory variables included in the choice models

Variables Questions asked in the survey Coding of variables Perceived threat of a) How likely is in your view a 0-20 scale constructed by a dam breach dam breach in Switzerland? summing up a) and b), both b) How strongly would you be measured on a 0-10 scale where affected by a dam breach? 0=’very unlikely/not at all’ and 10=’very likely/very much’ Perceived threat of a) How likely is in your view a 0-20 scale constructed by a nuclear accident nuclear accident in Switzerland? summing up a) and b), both b) How strongly would you be measured on a 0-10 scale where affected by a nuclear accident? 0=’very unlikely/not at all’ and 10=’very likely/very much’ Fear of a dam How concerned are you about a 0-10 scale where 0=’not at all’ breach dam breach? and 10=’very much’ Fear of a nuclear How concerned are you about a 0-10 scale where 0=’not at all’ accident nuclear accident? and 10=’very much’ Self-reported effect Has the risk ladder changed your 1=yes; 0=no of risk ladder on perception of hydropower risk? hydropower risk perception Self-reported effect Has the risk ladder changed your 1=yes; 0=no of risk ladder on perception of nuclear risk? nuclear risk perception Risk attitude Do you consider yourself, in 0-10 scale where 0=’not at all general, a person who takes risks willing to take risks’ and or do you prefer to avoid taking 10=’very willing to take risks’ risks? Endpoint taken Have you compared the risks of 1=yes; 0=no into consideration dying from a dam breach and a nuclear accident with the risk depicted at the endpoint of the risk ladder? Middle point taken Have you compared the risks of 1=yes; 0=no into consideration dying from a dam breach and a nuclear accident with the risk depicted in the middle of the risk ladder? Member of an Do you or anyone in your 1=yes; 0=no environmental household ﬁnancially support an organization environmental organization? Income What is your household’s gross 1=higher than the sample annual income? average; 0=sample average or lower Female What is your gender? 1=female; 0=male Age When were you born? 15-84 5.4. Econometric models and testing procedures 115 of choice tasks were designed. These were randomly assigned across respondents. Each respondent answered seven choice tasks. The ﬁnal DCE was admin- istered in June 2016 to a representative sample of 495 respondents belonging to the German- and French-speaking Swiss population.

5.4 Econometric models and testing procedures

Choice models have their roots in Lancaster’s theory of consumer choice, which posits that utility from a good is derived from the good’s characteristics or attributes (Lancaster, 1966). Based on this theory, the utility of an alternative in a choice model is deﬁned in terms of its attributes included in the choice tasks (and other variables). The usual econometric speciﬁcation follows McFadden’s random utility theory (McFadden, 1974), which assumes informational asymmetry between the respondent and analyst. More formally, the utility derived by respondent n from alternative i in choice situation t is:

Unit = Vnit + "nit, (5.3) where the error term "nit is unknown to the analyst. The unknown error term is responsible for the difference between the respondent’s actual utility Unit and the utility Vnit observed by the analyst. Distributional assumptions about the error term drive the distinction between the different types of models. For our analysis, we use the most commonly assumed iid extreme value distribution. The standard probability representation of the multinomial logit (MNL) is obtained by deﬁning a probabilistic condition for an individual n to choose alternative i over alternative j and combining this condition with equation 5.3. The result is shown in equation 5.4: x e 0n nit Pnit()= , (5.4) nxnjt j e 0 where the observed utility Vnit of equationP 5.3 is replaced by a linear speciﬁcation of a vector of parameters and observed variables xnit. We estimate mixed logit (MXL) models, which relax the restrictive assumptions that unobservable factors are uncorrelated across alternatives and choice tasks, and allow for taste heterogeneity between respondents:

x e 0n nit Pnit()= f()d. (5.5) nxnjt j e 0 Z ⇣ ⌘ P 116 Chapter 5. Reference Points for the Valuation of Risk Changes

The density f() can follow any distribution specified by the analyst. We assume f() to follow a normal distribution, as this can be motivated by the central limit theorem (Hensher and Greene, 2002). Estimation of equation 5.5 requires simulation. For a predefined number of draws, a value of is estimated, and a choice probability is calculated. The average of these probabilities is then taken to estimate equation 5.5 (Hensher, Rose, and Greene, 2015; Train, 2009). Finally, the MXL model is estimated employing maximum likelihood estimation on the simulated log-likelihood function. We apply the test procedure suggested by Swait and Louviere (1993) to test our first hypothesis. The Swait-Louviere procedure tests for equality of the entire vector of preference parameters between our two samples. To assess the second hypothesis: whether the different risk ladders have an effect on welfare estimates for a change in risk, we use the Poe procedure (Poe, Girarud, and Loomis, 2005) to test for differences between MWTP values for a change in hydropower and nuclear power risk derived from the two samples (see Appendix 5.A for details).

5.5 Results

5.5.1 Descriptive statistics

Both samples are drawn from the German- and French speaking population in Switzerland, accounting for roughly 95% of the total Swiss population. 25 respondents who chose the status quo option in all choice tasks were identiﬁed as protest responses (e.g. Brouwer and Martín-Ortega, 2012). After excluding these respondents from the analysis, the remaining two samples consist of 470 individuals in total. The survey response rate is 16.3%. Table 5.3 compares the characteristics of the two samples and the target population from which they were drawn. The sociodemographic characteristics of samples 1 and 2 are very similar and representative of the general population due to the applied sampling strategy, which ensured representativeness and similarity with respect to gender, age, linguistic background, and education. There are only minor differences between the two samples and the population from which the samples were drawn, such as the share of students and the share of respondents holding a university degree, which are slightly higher in the two samples compared with the target population, and the slightly lower unemployment rates in the two samples. The samples’ general risk attitudes and hydropower and nuclear power risk perception are presented in Figure 5.5. Figure 5.5 provides two main insights into 5.5. Results 117

TABLE 5.3: Sociodemographic characteristics of the study samples and target population

Sample 1 Sample 2 Populationa (High reference (Low reference point) point) Female (%) 53.2 49.8 50.5 Average age 47.2 47.9 41.5 French-speaking origin (%) 22.6 25.5 23.0 Average household size 2.5 2.5 2.3 University degree (Bachelor’s, 34.5 37.4 27.1 Master’s, PhD) (%) Average annual gross household 98,174b 101,523b 120,624 income (CHF) Unemployed (%) 1.7 2.1 4.3 Retired (%) 20.0 23.4 17.9 Student (%) 10.6 7.7 5.7 N 235 235 7,887,303 Notes: aSFOE (2016). bThis is an approximation based on midpoint estimates as respondents were asked to indicate in which income group their gross household income falls.

8 Sample 1 (high risk endpoint) Sample 2 (low risk endpoint)

0 Willingness to Perceived Perceived Affected by a Affected by a Worried about a Worried about a take risks likelihood of a likelihood of a dam breach nuclear accident dam breach nuclear accident dam breach nuclear accident

FIGURE 5.5: Descriptive statistics of risk attitudes and perceptions, from 0 (very unlikely / not at all) to 10 (very likely / very much)4 118 Chapter 5. Reference Points for the Valuation of Risk Changes public perception of hydropower and nuclear risk. First, the differences between sample 1 and sample 2 seem to be small. The non-parametric Mann-Whitney test cannot reject the equality of all variables between the two samples at the 10% signiﬁcance level, except for the perceived threat severity of a dam breach (affected by a dam breach). Second, the perceived threat severity, the perceived probability of threat occurrence (likelihood of a dam breach/nuclear accident), and the fear of an accident (worried about a dam breach/nuclear accident) are signiﬁcantly higher for nuclear power than for hydropower. Respondents’ perceived controllability of the comparative risks on the risk ladders was measured on a 5-point Likert scale, and are presented in Table 5.4.

TABLE 5.4: Perceived controllability of the comparative endpoint and middle point risks

To what extent do you agree with the following statement? ‘There is nothing I can personally do to reduce the following risks’ Endpoint Endpoint Middle point risk Middle point risk risk risk Sample 1 Sample 2 Sample 1 Sample 2 (cancer) (lightning (lightning strike) (severe strike) earthquake) Strongly disagree 20.9 21.7 14.9 3.0 (%) Disagree (%) 44.7 40.0 33.6 8.9 Neither agree nor 15.7 9.4 15.3 14.5 disagree (%) Agree (%) 14.0 17.9 21.7 23.4 Strongly agree (%) 4.7 11.1 14.5 50.2

Table 5.4 compares the perceived controllability of the endpoint and the middle point risks displayed on the risk ladders. Although the two endpoint risks differ with respect to their type and probability, the controllability of the two hazards is perceived to be similar: 65.6% and 61.7% of the respondents disagree or strongly disagree with the statement that they cannot do anything to reduce the risk of dying of cancer and dying from a lightning strike in sample 1 and 2, respectively. The share of participants who agree or completely agree with the statement is slightly higher for the risk of a lightning strike (29.0%) than for the risk of cancer (18.7%). More respondents neither agree nor disagree with the statement for the latter risk. 4Error bars reﬂect one standard deviation. 5.5. Results 119

A somewhat different pattern emerges when comparing the middle point risks of the two risk ladders. More respondents agree or completely agree that they can inﬂuence the risk of dying from a severe earthquake than the risk of dying from a lightning strike. The opposite result holds for participants who disagree or strongly disagree with the statement, whereas the share of respondents who neither agree nor disagree is similar for both risks. In conclusion, these results suggest reasonable comparability of the risks shown to the two samples, although the endpoint risks seem to be more comparable than the middle point risks. After displaying one of the two risk ladders, we asked the survey respondents whether the ladder had changed their perception of hydropower and nuclear power risks. The majority of respondents in both samples stated that the risk ladder had no effect on their risk perception. A similar share of respondents in samples 1 and 2 reported a change in nuclear risk perception induced by the risk ladders: 24.3% and 22.1%, respectively. However, the risk ladder with a wide range of comparative risks (high risk endpoint) had a somewhat stronger effect on hydropower risk perception (21.7% of the respondents report a change in risk perception) than the risk ladder with low comparative risk probabilities (14.9% report a change).

5.5.2 Choice model results

The results of the estimated MXL models are shown in Table 5.5. The second and third column report the results estimated for sample 1, and the fourth and ﬁfth column for sample 2. Both models are estimated using 5,000 Halton draws.

TABLE 5.5: Estimated mixed logit models

Sample 1: Sample 2: Risk ladder with a Risk ladder with a high reference point low reference point Variables Coeff. s.e. Coeff. s.e.

Mean estimates of random and non-random parameters Alternative-speciﬁc constant (ASC) -2.701** 1.162 -0.483 1.477 Construction of new hydropower plants -0.446*** 0.156 -0.540*** 0.157 Increase in risk of dying from a dam breach -0.928*** 0.282 -1.172*** 0.259 by 40% Decrease in risk of dying from a nuclear 0.531** 0.252 1.127*** 0.287 accident by 60% Increase in annual electricity bill (in 100 CHF) -0.408*** 0.047 -0.354*** 0.049 120 Chapter 5. Reference Points for the Valuation of Risk Changes

Sample 1: Sample 2: Risk ladder with a Risk ladder with a high reference point low reference point Variables Coeff. s.e. Coeff. s.e.

Increase in annual electricity bill Income 0.007 0.061 -0.035 0.065 ⇤ above sample average Perceived threat of a dam breach (before the -0.112* 0.067 -0.161* 0.084 risk ladder) Perceived threat of a nuclear accident (before -0.016 0.077 0.142* 0.086 the risk ladder) How concerned are you about a dam breach? -0.235* 0.129 -0.231 0.184 (before the risk ladder) How concerned are you about a nuclear 0.194 0.119 0.304** 0.142 accident? (before the risk ladder) Has the risk ladder changed your perception 1.631*** 0.627 0.032 0.846 of the risk of a dam breach? (1=yes, 0=no) Has the risk ladder changed your perception -1.158** 0.566 -0.260 0.732 of the risk of a nuclear accident? (1=yes, 0=no) Endpoint risk taken into consideration 0.284 0.574 0.199 0.967 (1=yes, 0=no) Middle point risk taken into consideration 0.817 1.041 -0.573 0.915 (1=yes, 0=no) Risk attitude 0.120 0.101 0.095 0.126 Member of an environmental organization 1.243*** 0.472 2.416*** 0.635 (1=yes, 0=no) Member of an environmental organization -0.563** 0.231 -0.728*** 0.238 ⇤ Construction of new hydropower plants Female (1=yes, 0=no) -0.638 0.477 -1.469** 0.605 Age 0.012 0.014 0.029 0.018 Age Increase in risk of dying from a dam 0.013** 0.006 0.019*** 0.005 ⇤ breach by 40% Age Decrease in risk of dying from a 0.004 0.005 -0.008 0.005 ⇤ nuclear accident by 60% Standard deviations of random parameters ASC 2.434*** 0.285 3.186*** 0.357 Construction of new hydropower plants 1.298*** 0.140 1.292*** 0.144 Increase of risk of dying from a dam breach 0.847*** 0.130 0.555*** 0.154 by 40% Decrease of risk of dying from a nuclear 0.567*** 0.142 0.811*** 0.137 accident by 60% Model characteristics Number of observations 1645 1645 Log-likelihood (restricted) -1807.217 -1807.217 5.5. Results 121

Sample 1: Sample 2: Risk ladder with a Risk ladder with a high reference point low reference point Variables Coeff. s.e. Coeff. s.e. Log-likelihood (unrestricted) -1259.635 -1202.615 AIC/N 1.562 1.493 BIC/N 1.644 1.575 McFadden’s pseudo R2 0.303 0.335 Notes: ***p<0.01, **p<0.05, *p<0.1. AIC: Akaike Information Criteria; BIC: Bayesian Information Criteria.

The two choice models perform generally well according to the AIC, BIC, and pseudo-R2 statistics, although the model based on sample 2 has a slightly better fit. The standard deviations of the random parameters are highly significant in both models, suggesting that preferences for the non-monetary choice attributes vary across individual respondents. Many of the coefficients are similar between the two samples, and reveal common drivers of preferences for the hydropower expansion scenarios. All attributes have the expected signs and are highly significant. Compared with extending existing plants, the construction of new hydropower plants decreases the likelihood of choosing a hydropower expansion scenario, since the construction of new facilities has a stronger detrimental environmental effect than the extension of existing facilities. An increase in the risk of dying from a dam breach by 40% compared with an increase of 20% has a negative effect on the choice probability, whereas a decrease in the risk of dying due to a nuclear accident by 60% compared with 30% increases the probability of choosing a hydropower expansion scenario. Although the price attribute is, as expected, negative and statistically significant, its interaction with a binary variable that equals 1 if a respondent’s income is above the sample average is not statistically significance. In other words, the price sensitivity does not seem to depend on a respondent’s income, i.e. the choices of respondents with a high income are not significantly differently affected by an increase in the annual electricity bill than choices of respondents with an average or lower than average income. Although the signs of the significant coefficients are as expected, the coefficients for a perceived threat of hydropower and nuclear risks are only weakly significant, or insignificant in the case of a perceived threat of a nuclear accident in sample 1. They are negative if they relate to hydropower (risk increase) and positive if they relate to nuclear power (risk reduction). Respondents who perceive a dam breach (nuclear power accident) as a high threat are hence less likely (more likely) to choose a 122 Chapter 5. Reference Points for the Valuation of Risk Changes hydropower expansion scenario because it involves an increase (decrease) in hydropower (nuclear power) risk. The same logic holds for the coefficients for how concerned respondents are about a dam breach or a nuclear accident, although this effect is not significant in either sample. Contrary to the perceived risks, general risk attitudes have no significant impact on respondents’ choices in the two samples. The higher importance of risk perception than risk attitude has also been reported in previous studies (e.g. Slovic and Weber, 2011; Weber and Milliman, 1997). In order to control for the effect of reference points other than the status quo, we included two binary variables that indicate whether respondents considered the endpoint or middle point risk shown on the risk ladder in their choice process. Neither of these variables has a significant effect on stated choices. The sociodemographic characteristics reveal that female respondents in sample 2 prefer the status quo to the expansion scenarios. Note that this effect per- sists even when gender-specific risk effects are controlled for by including interaction terms between gender and the hydropower and nuclear risk attributes, and while controlling for interaction effects between gender and the self-reported effect of the risk ladder on respondents’ risk perceptions. All of these interaction terms never reach significance and are not included in the final models. Age interacted with the increase in the risk of a dam breach results in a significant positive coefficient, indicating that older respondents are less concerned about an increase in hydropower risk than younger respondents, possibly reflecting a higher level of trust in hydropower by older generations. Members of an environmental organization have a higher preference for hydropower expansion scenarios than for the status quo. This holds even if an interaction term between membership of an environmental organization and the construction of new hydropower plants (involving stronger negative environmental effects) is included. This latter interaction term is, as expected, significant and negative. Turning to the main differences between the two samples, first, the ASC is highly significant and negative for sample 1, but not significant for sample 2. This indicates that respondents in sample 2 are indifferent between the expansion scenarios and the status quo, and made their choice exclusively on the basis of the presented attributes. In contrast, since the ASC was part of the status quo utility specification, respondents in sample 1 prefer an expansion of hydropower over the status quo. This may be an indication for the status quo to be of importance as a reference point in sample 1 but not in sample 2. Second, with the exception of the price attribute, the magnitudes of the attribute coefficients are 5.5. Results 123 higher for sample 2 than for sample 1, suggesting a choice process that is more strongly influenced by the attributes. Nevertheless, the coefficients and their standard errors overlap with the exception of the nuclear risk attribute. Finally, the coefficients for the variables indicating whether the risk ladder has changed respondents’ risk perception are only significant in sample 1. This means that the impact of the risk ladder on risk perception is more relevant for the choice process of respondents in sample 1 than in sample 2, although the difference in the share of respondents who reported a change in risk perception due to the risk ladders is small between the two samples and negligible in the case of nuclear risk (see Section 5.5.1). The change in risk perception induced by the risk ladders has a positive effect on the probability of choosing an expansion scenario in the case of hydropower risk and a negative effect in the case of nuclear risk. This is most likely because a higher share of respondents stated that the risk ladder decreased their risk perception than those who stated the opposite. This holds for both hydropower and nuclear risks. That is, more respondents seemed to initially overestimate the actual risk levels. As a result, if the risk ladder decreased the perception of hydropower risk, a positive effect on choosing an expansion scenario is expected. The opposite effect is expected for nuclear power risk.

5.5.3 Hypotheses test results

Our main hypothesis states that a change in risks depicted on a risk ladder with a narrow range of comparative risks (sample 2) has a larger effect on utility and choices than a change in risks shown on a risk ladder with a wide range of comparative risks (sample 1). As a first test for the equality of the parameters between the two samples, we ran the Swait-Louviere test procedure. The results show that neither the equality of preference parameters nor the equality of scale parameters between samples 1 and 2 can be rejected (the outcome of the LR test statistic is 28.27 and 0.51 with corresponding p-values of 0.35 and 0.47 for the first and second test stage, respectively). As mentioned, the magnitude of the attribute coefficients are higher for both risks in sample 2 than in sample 1, albeit significantly larger only for the case of nuclear risk. A more formal test of our hypothesis is based on the comparison of MWTP values, since we expect the MWTP for changes in nuclear and hydropower risk to be significantly higher for sample 2 than sample 1. The MWTP values estimated in preference-space based on the choice models are reported 124 Chapter 5. Reference Points for the Valuation of Risk Changes in Table 5.6. The MWTP values correspond to an increase in an average household’s annual electricity bill of between 14% and 34%.

TABLE 5.6: MWTP estimates for the risk attributes (CHF per household per year)

Sample 1: Sample 2: High risk reference Low risk reference point ladder point ladder Choice attributes MWTP s.e. MWTP s.e. Increase if the risk of dying from a dam 227.33*** 73.95 313.39*** 86.80 breach by 20% relative to 40% Decrease in the risk of dying from a nuclear 130.01** 65.42 318.62*** 96.63 accident by 60% relative to 30% Notes: ***p<0.01, **p<0.05, *p<0.1. Standard errors are based on the Krinsky and Robb (1986) method using 10,000 draws.

At first sight, Table 5.6 supports our hypothesis, since the MWTP values for the risk attributes in sample 2 are consistently higher than those of sample 1. To test whether the differences are statistically significant, we apply the complete combinatorial method suggested by Poe, Girarud, and Loomis (2005). The outcome of this test indicates that our null hypothesis of equality of the MWTP values between the two samples can be rejected for the nuclear risk attribute (p=0.061) but not for the hydropower risk attribute (p=0.239). Therefore, the two different risk ladders only seem to have a significantly different effect on MWTP for a change in nuclear risk.

5.6 Discussion and conclusions

The results obtained in this study are partly in line with our theoretical expectations of comparative risk probabilities in risk ladders to serve as reference points. Specifically, we expect the effect of a risk ladder with a wide range of comparative risks (sample 1) to have a smaller effect on respondents’ choice than a risk ladder with a narrow range of comparative risks (sample 2). This is confirmed by the MWTP values for identical changes in risk which are smaller in sample 1 than in sample 2. However, the differences in estimated coefficients and MWTP values between the two samples are only significant for the nuclear risk attribute. The estimated choice models suggest a possible explanation for this outcome: 5.6. Discussion and conclusions 125 the reduction in risk perception induced by the risk ladder with high comparative risks affects respondents’ choices more strongly, and hence results in lower MWTP values than the reduction in risk perception induced by the risk ladder with low comparative risks. The Swait-Louviere test is not able to reject the null hypothesis of equality of preference parameters between the two samples, but this is considered a weaker test in view of the fact that this test refers to differences in the full set of preference coefficient estimates. A further interesting finding is that the ASC is highly significant in the model based on sample 1 but does not reach significance in the model based on sample 2. This suggests that, at least for sample 1 and its high comparative risks, the status quo also serves as a reference point. Note that this result needs to be interpreted with care for two reasons: First, the coefficient for the ASC reflects the impact of the status quo levels of all attributes included and not only of the risk-related attributes; and, second, it may be confounded with the effect of omitted variables not captured by the choice attributes and other explanatory factors. In conclusion, our null hypothesis of equal MWTP values irrespective of the relative risks displayed on the risk ladder is rejected for the change in nuclear risk, but not for the change in hydropower risk. At the same time, and at least for the sample with the wide range of comparative risks, it is unlikely that the comparative risks on the risk ladders serve as unique reference points. The conclusions that can be drawn from this study highlight the prevalence of multiple reference points in DCEs besides the status quo. This contradicts part of the existing literature that typically focuses on a single reference point. We find empirical support for the untested expectation in Sandman, Weinstein, and Miller (1994) that not only the location of the risk that is valued matters, but also the range of values on the risk ladder. We are unable to conclude whether a risk ladder with comparative risks that are more or less similar in probability to the risk in question performs better in terms of risk communication, but our results indicate that comparative risks which are closer in probability to the risks in question result in higher MWTP values. The implications of our main findings for the valuation literature which investigates reference points highlight the need for a broader perspective. Con- trary to the common notion that design elements of DCEs serve as reference points and introduce procedural bias, we show that there may be additional elements related to the accompanying survey that also serve as reference points, in casu, the role of risk ladder design. However, the importance of comparative risks as reference points may also be valid for situations involving other, 126 Chapter 5. Reference Points for the Valuation of Risk Changes graphical or textual, risk communication devices. Hence, a careful construction and pretesting of such devices seems necessary. More research is needed to determine the relative importance of the different possible reference points in risk communication devices. A promising path for future research may be to com- bine research on the effects of risk communication device specifications with a tool to monitor the visual information acquisition process of respondents, such as eye-tracking or mouse-tracking. 5.A. Poe-test 127

5.A Poe-test

The complete combinatorial approach suggested by Poe, Girarud, and Loomis (2005) is based on the more complex convolution approach (Ohdoko, 2008; Poe, Severance-lossin, and Welsh, 1994), but similarly provides an unbiased assessment of the statistical signiﬁcance of differences in two simulated distributions. In order to calculate the Poe-test statistic, we start with simulating a distribution of the MWTP values for a change in risk for sample 1 and sample 2, denoted by Xˆ and Yˆ , respectively. Both distributions are based on 1000 random and normally distributed draws with the means and standard deviations of the estimated MWTP values. The null hypothesis that is to be tested is the following:

H : Xˆ Yˆ =0. (5.6) 0

The test statistic for H0 is the empirical cumulative distribution function at zero:

Yˆ = #(x y 0) /(m n), (5.7) i j  ⇤ with # denoting the number of differences that meet the contained condition and m and n to equal the number of draws of Xˆ and Yˆ ; respectively (Ohdoko, 2008; Poe, Girarud, and Loomis, 2005). The test statistic in equation 5.7 functions identically as a p-value: It indicates the probability that expresses the rejection area of the null hypothesis (Ohdoko, 2008).

129

Chapter 6

Conclusions, policy recommendations, and future research directions

6.1 Summary of the main ﬁndings

The nuclear accident of 2011 in Fukushima, Japan, and the subsequent transition of the world’s energy supply towards the increasing use of renewable sources spurred the Swiss government to develop a new long-term energy policy (SFOE, 2013). A key pillar of this policy, called "Energy Strategy 2050", is the expansion of hydropower. Taking into account that the hydropower industry in Switzer- land faces considerable uncertainty with respect to its long-term economic vi- ability and its public acceptance, an expansion of hydropower production is a challenging task. Embedded in this policy context, this dissertation has aimed to shed light on the public’s preferences for hydropower, building upon international experience by means of a global meta-analysis, and by a discrete choice experiment (DCE) carried out among a representative sample of the Swiss population. The meta- analysis quantitatively analyzed the economic values for hydropower externalities, and explained their determinants based on available secondary data from the existing literature, while the DCE elicited public preferences and willingness- to-pay (WTP) for a hypothetical expansion of hydropower among the Swiss population. The central hypothesis postulated in this PhD thesis is that public preferences for an expansion of hydropower in Switzerland are intrinsically linked 130 Chapter 6. Conclusions to the phasing-out of nuclear power, as the Swiss Energy Strategy 2050 envis- ages a partial replacement of nuclear energy with hydropower. An expansion of hydropower is likely to render the phasing-out of nuclear power more probable, because it would compensate for a decline in the country’s electricity production. This PhD dissertation has contributed to the existing literature by demonstrating the importance of nuclear risk as an indirect externality of hydropower. The role of nuclear risk perception in public preferences for hydropower expansion was captured by including nuclear risk as an explicit choice attribute in the DCE design. This is the first study in the stated preference (SP) literature that explicitly links public preferences for reducing nuclear risk to public preferences for a renewable source of energy. On the basis of the findings, it is recommended that future SP studies which elicit preferences for a specific source of energy should also take into account public preferences, attitudes, and the WTP for other sources of energy. The main focus of the chapters of this thesis was to test two standard economic axioms underlying consumer choice (monotonicity and continuity), as well as the conventional assumption that preferences are known and stable in the context of a DCE. Moreover, the significance of prospect theory, and, in particular, the reference dependence of choice behavior, in this specific risk management context was examined. Chapter 2 aimed at explaining the variation observed in the non-market values for hydropower externalities worldwide. For this purpose, a quantitative meta-analysis of existing studies that estimated the non-market values of hydropower externalities was performed. The contribution presented here is novel, since this is the first quantitative meta-analysis that focuses exclusively on hydropower externalities. Sample and study characteristics were controlled for in the estimated meta-regression models. The obtained results suggest the existence of public aversion towards the deterioration of landscape, vegetation, and wildlife caused by hydropower operations. In contrast, empirical evidence on respondents’ WTP for mitigating the hydropower impacts on these environmental resources is limited. The avoidance of greenhouse gas emissions, which is the most important positive external effect of hydropower electricity generation, proved to exert a significant positive influence on welfare estimates, but only in combination with the share of hydropower in a country’s national electricity production. In other words, the positive effect of hydropower on greenhouse gas emissions is only valued in countries that already have a high share of hydropower in their national electricity supply. This is possibly related to 6.1. Summary of the main findings 131 a higher level of public awareness in these countries of the beneficial effects of hydropower on greenhouse gas emissions. It was also found that the impacts of hydropower on aesthetic and recreational amenities do not exert a significant influence on welfare measures. The analysis furthermore showed sensitivity to scope across all externalities. The results from the meta-analysis served as inputs in designing the DCE on a hypothetical expansion of hydropower in Switzerland. The DCE aimed to elicit public preferences for the proposed hydropower extension, accounting for the reduction in nuclear power risk, and at the same time answer the main methodological research questions of this dissertation. Chapter 3 investigated the common and idiosyncratic determinants of choice certainty, choice consistency, and choice monotonicity in DCEs. These factors are linked to the axiom of monotonicity and the standard assumptions of known and stable preferences made in microeconomic theory. Adding to the existing literature, Chapter 3 investigated these concepts simultaneously and was based on the same sample of respondents. This allowed for an assessment of their common and idiosyncratic determinants. The results provided several insights. First, there are significant differences between the choice behavior of certain and uncertain respondents, as well as between consistent and inconsistent respondents. Second, no procedural effect of posing the choice certainty question was found. This holds for both an entire choice-task sequence and individual choice tasks. The latter effect had not been tested before in the literature. Third, the position of a repeated choice task had no effect on choice consistency. Finally, a variety of idiosyncratic determinants of choice certainty, consistency, and monotonicity were identified, and only gender and the utility difference between choice alternatives were identified as common drivers. It was found that female respondents display less certainty but a higher degree of consistency and monotonicity in their choice behavior. Both measures of choice-task complexity were found to be relevant for choice certainty, but only the utility difference was reported to have an effect on consistency and monotonicity, by and large consistent with previous research in this particular area of choice-task complexity. A common violation of the standard economic axiom of preference continuity in DCEs is attribute non-attendance (ANA). Chapter 4 aimed to answer the research question which concerned how a visual measure of ANA based on mouse-tracking performs in explaining ANA behavior compared with stated and inferred ANA. It represents the first study that investigates ANA in DCEs by 132 Chapter 6. Conclusions means of mouse-tracking. Contrary to eye-tracking, mouse-tracking records visual activity online, which has a number of distinct advantages over eye-tracking, such as lower survey costs and access to a larger pool of potential respondents. The results of Chapter 4 support the findings reported in the existing studies that analyze ANA using eye-tracking data (Balcombe, Fraser, and McSorley, 2015; Van Loo et al., 2014). The performance of the choice models that incorporate mouse-tracking information on ANA was compared with models that use stated and inferred ANA information. The results suggest that choice models estimated using visual ANA data do not outperform the models estimated using stated ANA information. Nevertheless, choice models based on visual ANA result in a slight improvement over both choice models that do not take ANA into account and choice models that use inferred ANA information. Finally, Chapter 5 addressed the question whether comparative risks shown on risk ladders serve as reference points, and hence influence the preferences and welfare estimates for changes in risks. Chapter 5 is concerned with a key assumption in prospect theory, i.e. the dependence of preferences on reference points. As opposed to most of the valuation literature, which assumes that reference points coincide with the status quo, this chapter hypothesized that different ranges of probabilities of comparative risks on risk ladders also serve as reference points in the choice process of survey respondents. The obtained results support theoretical expectations: While keeping the location of the valued risk change constant in the two risk ladders, the ladder with a wide range of probabilities associated with comparative risks has a smaller impact on respondents’ choices than the risk ladder with a narrow range of probabilities. This finding is supported by the observed differences in marginal WTP for the risk changes between the two samples that were shown the two different risk ladders. More generally, the findings in Chapter 5 suggest that it is improbable that comparative risks serve as a unique reference point, and it is expected that multiple reference points, including the status quo, influence respondents’ choices in DCEs.

6.2 Directions for future research

This dissertation suggests a number of new directions for future research concerning the methodological aspects of DCEs. To start, the work presented here emphasizes the potential benefits of using visual tracking technologies to better monitor and understand the visual information acquisition process of DCE 6.2. Directions for future research 133 respondents. Although the impact of using mouse-tracking information in the estimated choice models in Chapter 4 proved to be limited, more studies using mouse-tracking are needed to assess the robustness of the results found here, as it provides several crucial advantages over eye-tracking. In general, mouse- tracking can be used to extend the limits to the amount of information that can be inferred from stated choice data, since deeper insights into the axioms and assumptions of consumer choice theory can be gained by establishing a clearer relationship between respondents’ visual activity and choice behavior. Specific research questions of interest here include, among others, "How does the visual activity of certain, consistent, and monotonic respondents in a DCE vary between uncertain, inconsistent, and non-monotonic participants?" and "How does the understandability of a DCE or task complexity relate to their visual representation?" The visual activity of respondents who state a high survey understandability, or who spend much time on informational pages, may be contrasted with the visual activity patterns of survey participants who state a low understandability or who spend only a little time on informational pages. Such research may provide detailed insights into the characteristics of a well-understood DCE. Also, the tracking of the visual information acquisition process may contribute to understanding gender differences related to the choice-making process, which is currently an under-researched aspect. Such research could lead to improvements in the design of DCEs, and produce results that are less prone to choice uncertainty, inconsistency, and non-monotonicity, hence increasing the conformity of choice behavior with the assumptions underlying consumer choice in economics. Research using visual data could provide more definite insights into precisely how ANA behavior manifests itself visually, and increase the reliability of the results found in this PhD thesis. Further studies could possibly aim to ‘dynam- ically’ adjust a choice design based on an individual’s visual activity during a choice-task sequence. It could, for example, be investigated whether adjusting the position of an attribute in a choice task alters the visual attendance by respondents who displayed non-attendance to this attribute in prior choice tasks. Similarly to how mouse-tracking can be used to gain information on ANA, there is potential to use mouse-tracking for investigating other heuristics in attribute processing, such as lexicographic or elimination-by-aspects decision strategies. As an example, further research could focus on whether common-metric attributes are evaluated separately or combined. Frequent gaze shifting between two common-metric attributes as revealed by mouse-tracking may indicate a 134 Chapter 6. Conclusions combined evaluation of attributes. Finally, visual data can also provide more in-depth knowledge about the premises of prospect theory. This thesis argued for the existence of reference points that are associated with the communication of risk information preceding the actual choice tasks. Visual data could help to identify relevant reference points by mapping the visual uptake of information to choice behavior. As an example, the relative importance of textual and graphical information and their potential role as reference points could be further investigated. However, visual tracking methods also seem to have a limit with respect to how much information on choice behavior can be obtained. Deeper insights may be gained by applying neuroeconomic methods. Neuroeconomics is the logical continuation in the development of economic theory. While behavioral economics attempts to uncover the psychological basis of economic choice in contrast to standard economic theory, neuroeconomics aims to explain economic behavior from a biological point of view. Neuroeconomic studies assess the neural antecedents of economic behavior, and try to find neural representations of value and utility (see, e.g., Bruce, Crespi, and Lusk (2015) and Lusk et al. (2016) for and overview). Up till now, there have been only a few applications of neuroeconomics in the realm of the economic valuation of environmental resources (e.g. Sawe, 2017; Sawe and Knutson, 2015). Most neuroeconomic valuation studies use functional magnetic resonance imaging (fMRI), when respondents answer valuation questions in a DCE or a contingent valuation setting. fMRI tracks blood oxygenation level-dependent signals over time which represent neural activity in a certain brain region (Sawe, 2017). Some studies that investigate the link between the activity in certain parts of the brain and choice behavior have been able to predict individual choice decisions (Lusk et al., 2016). Other studies have managed to: distinguish true zero values from protest bids; measure the degree of abstraction that a scenario has for a survey respondent; visualize the influence of positive or negative affective responses on WTP; and link neural responses associated with specific choice attributes to ANA behavior (Sawe, 2017; Sawe and Knutson, 2015). In conclusion, more research on the neural correlates of environmental choice behavior would likely yield more profound insights into the heuristics and biases that characterize the behavior of respondents in DCEs. 6.3. Policy recommendations 135

6.3 Policy recommendations

The results obtained from the meta-analysis in Chapter 2 provide several policy- relevant indications. A strong impact of negative hydropower externalities on economic values, combined with a lack of significant public WTP for mitigating such effects, suggests a challenging point of departure for policies which aim to expand hydropower and/or mitigate its negative environmental impacts. This result shows that the prevention of negative environmental externalities of hydropower projects, such as the impacts on landscape, vegetation, and wildlife, is paramount. Projects in environmentally sensitive areas, for example, conservation areas or national parks, are hence likely to meet public resistance. At the same time, mitigating the negative environmental effects of a hydropower project should not lead to additional public spending. The results of the meta- analysis also suggest that an expansion of hydropower is more likely to receive public support in those countries that already have a high share of hydropower in their national electricity production. Finally, and in contrast to the findings of the literature on wind power (e.g. Mattmann, Logar, and Brouwer, 2016b) and photovoltaics (e.g. Faiers and Neame, 2006), the aesthetic and recreational impacts of hydropower projects do not prove to play a significant role and, hence, represent secondary concerns in planning procedures. In this dissertation, the DCE conducted on a representative sample of Swiss respondents produced various results that are relevant for Swiss energy policy. First of all, an initial survey question on a hypothetical popular vote indicates that an overwhelming majority of the survey participants (92%) stated that they would vote in favor of an expansion of hydropower if they had the opportunity to do so. Moreover, 78% of the survey respondents expressed their approval of phasing-out nuclear power. This finding is corroborated by respondents’ answers about preferred energy sources: Only 5.7% of the respondents included nuclear power among their preferred sources of energy, while 63.3% of them claimed that hydropower would represent one of their favorite sources of energy. Only coal and oil are less preferred primary sources of energy than nuclear power, whereas only solar and wind power are preferred to hydropower. The importance of the choice attribute concerning a reduction in nuclear risk for the preferences for the potential expansion of hydropower supports the need for a holistic view on energy policy. This is relevant not only for the purpose of research design, but also from a policy point of view. Changes in energy policy are more likely to receive public support if the anticipated relationships between 136 Chapter 6. Conclusions the different sources of energy are clearly explained. The choice model estimated for the entire sample of survey respondents indicates a high WTP for the extension of existing hydropower plants relative to the construction of new plants (roughly 180 Swiss Francs (CHF) per household and per year). This finding reflects substantial public WTP to avoid the detrimental environmental effects of additional hydropower generation, since respondents were informed that extensions lead to weaker, and new constructions to stronger, environmental consequences. This positive WTP for reducing the negative environmental effects of hydropower contradicts the results from the international literature synthesized in the meta-analysis. The Swiss public seems to be willing to pay a significant amount of money in order to avoid the adverse environmental impacts of hydropower, which is not unambiguously the case in a global setting. In Switzerland, an expansion of hydropower that focuses on the extension of existing plants is hence more likely to be successful than an expansion that requires the construction of new facilities, even if the former involves higher costs. Finally, the risks associated with hydropower expansion and the phasing- out of nuclear power are important drivers of public preferences. The annual average WTP per household for an increase in the risk of dying from a dam breach by 20% instead of 40% amounts to roughly 70 CHF, while the WTP for a reduction in the risk of dying from a nuclear accident by 60% instead of 30% equals 160 CHF. Although these money values seem high, they only imply an increase of the average annual electricity bill of a Swiss household of between 8% and 19%. Assuming an expansion scenario that resulted in the lowest WTP, i.e. an expansion based on constructing new plants that is associated with an increase in hydropower risk of 40% and a decrease in nuclear risk of 20%, the average annual WTP per household for an expansion of hydropower amounts to an increase of the current average electricity bill of almost 50%. In conclusion, these results suggest a considerable WTP for a future expansion of hydropower in Switzerland. 137

Appendix A

Choice Experiment Survey

[Note: Appendix A presents a translation of the baseline survey version that was shown to 250 respondents. There were three other versions shown to a total of 750 respondents (not shown here). Also, only the choice tasks that were part of one of the two blocks of choice sets is presented. The original version of the survey is in German and available from the author upon request.]

A.1 Thank you very much for choosing to take part in this survey

This survey is part of a research project that is conducted by ETH’s water research institute EAWAG in Zürich. The research project is ﬁnanced by EAWAG and serves only academic purposes. The survey aims at investigating public opinion about the future of hydropower and electricity supply in Switzerland. We are only interested in your opinion. There are no right or wrong answers. All your responses will be treated conﬁdentially and used solely for the purpose of this research project. Answering the questions will take about 30 minutes.

A.2 Your electricity consumption

1. Can you tell us how much your household pays about for electricity every year?

Yes. About how much? [. . . CHF/year] • 138 Appendix A. Choice Experiment Survey

No • 2. Do you switch off the lights when leaving a room at home?

Never • Rarely • Sometimes • Often • Always • 3. Do you pay attention to energy labels when buying household appliances such as a vacuum cleaner or lightbulbs?

Never • Rarely • Sometimes • Often • Always • 4. Do you keep your TV on stand-by mode if it is not in use?

Never • Rarely • Sometimes • Often • Always • I don’t have a TV • 5. Would you prefer a speciﬁc energy source of your electricity if you had a choice? (several answers are possible)

Yes • – Gas-powered electricity – Oil-powered electricity – Coal-powered electricity – Nuclear-powered electricity – Hydro-powered electricity A.3. Hydropower and nuclear power 139

– Solar-powered electricity – Wind-powered electricity – Geothermal-powered electricity (Electricity from the heat of the earth) – Biomass-powered electricity (Electricity from organic matter) – Other source(s): . . . No •

A.3 Hydropower and nuclear power

Please read the following information first before continuing with the next question. Electricity is predominantly produced by hydropower and nuclear power in Switzerland. There are two main types of hydropower plants: Storage power plants (with dams) and run-of-the-river plants (without dams). Hydropower is a renewable energy source because it does not deplete natural resources. Hydropower produces virtually no greenhouse gases or other air pollutants. Plants with dams however involve a small risk of dam breaches. Both types of hydropower plants may have negative effects on landscape, wildlife and vegetation Nuclear energy is not considered a renewable energy because it uses ura- nium, a metal that is abundant but not infinitely available. Nuclear energy produces virtually no greenhouse gases or other air pollutants. However, it involves a small risk of a nuclear accident. Nuclear energy also produces radioactive waste, which is difficult to dispose of and needs to be stored safely for a very long time Dam breaches and nuclear accidents are very unlikely to happen but can have catastrophic consequences. Dam breaches can result in flood waves, while nuclear accidents may release radiation. Both events may kill hundreds or even thousands of people.

A.4 Your opinion about dam breaches

1. How likely do you think is a dam breach in Switzerland? (0=very unlikely, 10=very likely)

0 10 140 Appendix A. Choice Experiment Survey

2. There are about 200 dams in Switzerland. The possible consequences of a dam breach depend on the location and size of the dam. Please estimate: What is the average estimated risk of dying from a dam breach in Switzer- land? (Please try to approximately estimate the question if you are unsure.)

One in . . . people run the risk of dying from a dam breach during their • lifetime.

3. How much do you think would a dam breach affect you and your closer family? (0=not at all, 10=very much)

0 10

4. How worried are you about dam breaches in Switzerland? (0=not at all worried, 10=very worried)

0 10

5. To what extent do you agree with the following statements? (0=don’t agree at all, 10=totally agree)

No dam breaches can happen because all risk factors are controlled • for. 0 10

The risk of dam breaches is being exaggerated. • 0 10

A.5 Your opinion about nuclear accidents

1. How likely do you think is a nuclear accident in Switzerland? (0=very unlikely, 10=very likely)

0 10

2. There are ﬁve nuclear power stations in Switzerland and two in bordering areas in France. The possible consequences of a nuclear accident depend on A.6. Risk graph 141

the location of the nuclear power station and the weather conditions at the time of an accident. Please estimate: What is the average estimated risk of dying from a nuclear accident in Switzerland? (Please try to approximately estimate the question if you are unsure.)

One in . . . people run the risk of dying from a nuclear accident during • their lifetime.

3. How much do you think a nuclear accident would affect you and your closer family? (0=not at all, 10=very much)

0 10

4. How worried are you about nuclear accidents in Switzerland? (0=not at all worried, 10=very worried)

0 10

5. To what extent do you agree with the following statements? (0=don’t agree at all, 10=totally agree)

No nuclear accidents can happen because all risk factors are controlled • for. 0 10

The risk of nuclear accidents is being exaggerated. • 0 10

A.6 Risk graph

Please have a look at the following graph. This graph compares the estimated current risk of death by a nuclear accident and a dam breach with other risks of death in Switzerland. All the risks shown describe the estimated average risk of dying by these causes in Switzerland during a lifetime. (Source: Estimates based on historical data by the federal statistical ofﬁce, Neuchatel, and the Paul Scherrer institute, Villigen) 142 Appendix A. Choice Experiment Survey

Risk of dying from 1 in 10 people 1 in 10 cancer

1 in 100

1 in 1000

1 in 10,000

1 in 35,000 people Risk of dying from a lightning strike

1 in 100,000

Risk of dying from 1 in 900,000 people a dam failure 1 in 1 Million

1 in 3 Mio. people Risk of dying from a nuclear accident

The estimated average risk of dying from a dam breach in Switzerland during a lifetime is currently 1 in 900,000. This means that 1 in 900,000 people risk dying from a dam breach. The estimated average risk of dying from a nuclear accident in Switzerland during a lifetime is currently 1 in 3M This means that 1 in 3M people risk dying from a nuclear accident Both estimates consider the immediate consequences of an accident as well as the associated long term effects.

1. Has this graph changed your perception of hydropower risk?

Yes. In what way? . . . • No • 2. Has this graph changed your perception of nuclear risk?

Yes. In what way? . . . • A.7. Expansion of hydropower 143

No • 3. To what extent do you agree with the following statement? There is nothing I can personally do to reduce the following risks.

Cancer • – Strongly disagree – Disagree – Neither agree nor disagree – Agree – Totally agree Lightning strike • – Strongly disagree – Disagree – Neither agree nor disagree – Agree – Totally agree Nuclear accident • – Strongly disagree – Disagree – Neither agree nor disagree – Agree – Totally agree Dam failure • – Strongly disagree – Disagree – Neither agree nor disagree – Agree – Totally agree

A.7 Expansion of hydropower

Please read the following information ﬁrst before continuing with the next question. 144 Appendix A. Choice Experiment Survey

Switzerland intends to phase-out nuclear energy within the next decades. The federal government plans to gradually replace the electricity produced by nuclear power with electricity from various renewable and non-renewable sources to meet future energy demand in Switzerland. As part of this plan, the government wants to increase the production of electricity by hydropower. At the moment, roughly 55% of electricity in Switzerland is produced by hydropower, whereas 40% stem from nuclear energy and 5% from other sources.

A.8 Your opinion in a public vote

1. If you could participate in a public vote on phasing-out nuclear energy in Switzerland, would you vote in favor of a phase-out or against this suggestion?

In favour • Against • I don’t know • 2. If you could participate in a public vote on expanding hydropower generation in Switzerland, would you vote in favor of an expansion or against this suggestion?

In favour • Against • I don’t know •

A.9 There are different possibilities how hydropower can be expanded

Please read the following information carefully before continuing with the next question. The federal government plants to expand the production of hydropower. An expansion of hydropower, however, can be done in several ways. In order to learn something about your opinion on the different ways of how hydropower can be expanded, we will show you eight decision situations (though experiments) with different options on the expansion of hydropower. A.10. How hydropower can be expanded (II/II) 145

We will request you to choose your preferred option. All choice situations will also include a "no expansion" option. In this case there will be no expansion of hydropower. First we will now show you some information on how the different options on expanding hydropower differ from each other. You will need the following information and explanations of symbols for answering the decisions situations afterwards.

A.10 Please read the following information about the different possibilities how hydropower can be expanded

An expansion of hydropower can be done by extending existing or by constructing new plants. Both options increase the share of hydropower in Swiss electricity production to the same extent. Constructing new plants has a stronger effect on the landscape. New plants also create additional threats to wildlife (e.g. for ﬁsh) and vegetation. These effects are less pronounced for the case of extending existing plants. The following options are possible:

• No expansion of hydropower • No additional effects on landscape, wildlife, and vegetation

• Extending existing plants • Little effects on landscape, wildlife, and vegetation

• Constructing new plants • Strong effects on landscape, wildlife, and vegetation

An expansion of hydropower increases the risk of dying from breaches. The following options are possible depending on whether the expansion is based on hydropower plants with or without dams: 146 Appendix A. Choice Experiment Survey

• No expansion of hydropower Current risk • 1 in 900,000 people risk dying from a dam breach in Switzerland

• The expansion is based on hydropower plants with and without dams • 1 in 750,000 people risk dying from a dam breach in Switzerland +20% risk

• The expansion is based on hydropower plants with dams +40% risk • 1 in 650,000 people risk dying from a dam breach in Switzerland

An expansion of hydropower reduces the risk of death by a nuclear accident. An expansion of hydropower facilitates the phase-out of nuclear power in Switzerland. This would lower the risk of a nuclear accident, although the risk of an accident in a nuclear power station in bordering regions of France remain. The following options are possible depending on how many nuclear power plants are switched off:

• No phase-out of nuclear energy Current risk • 1 in 3M people risk dying from a nuclear accident in Switzerland

• The three oldest nuclear power plants in Switzerland will be switched off • -30% risk 1 in 4M people risk dying from a nuclear accident in Switzerland

• All five nuclear power plants in Switzerland will be switched off • 1 in 7M people risk dying from a nuclear accident in bordering regions -60% risk of France

An expansion of hydropower increases your annual electricity bill. An expansion of hydropower costs money. The costs of such an expansion would need to be borne by electricity consumers. The yearly electricity costs of your household can therefore increase by the following amounts:

0 CHF No expansion of hydropower

From 100 to 600 CHF Expansion of hydropower A.11. Example situation 147

A.11 Here you see an example of a decision situation

Please have a look at the example situation. Subsequently, we will ask you to make a choice in eight such decision situations.

Optionen A) und B): Option C): Expansion of hydropower with effects on landscape, No expansion of hydropower wildlife and vegetation Risk of death by a dam breach as well The risk of death by a dam breach increases and the as a nuclear accident remain risk of death by a nuclear accident decreases unchanged An expansion of hydropower costs money Your electricity bill does not change either

A) Ausbau B) Ausbau C) Kein Ausbau

Type of hydropower New construction Extension No expansion expansion (Strong effects) (Little effects) (No additional effects)

Risk of dying from a +40% risk +20% risk Current risk dam breach (1 in 650,000 people) (1 in 750,000 people) (1 in 900,000 people)

Risk of dying from a -30% risk -60% risk Current risk nuclear accident (1 in 4M people) (1 in 7M people) (1 in 3M people)

Increase in your +100 CHF/year +300 CHF/year +0 CHF/year household’s yearly electricity bill

A.12 Eight hypothetical decision situations with different options on the expansion of hydropower will follow

In each of these choice situations we would like to know which of the options you prefer. Each choice should be made independently of previous choices. 148 Appendix A. Choice Experiment Survey

Please keep in mind that your income is constrained. The money you indicate to be willing to spend for an expansion of hydropower will not be available for the consumption of other goods anymore. Answer the questions as if you would actually need to pay the amount associated with your decisions.

A.13 Decision 1

1. Which option would you prefer?

A) Expansion B) Expansion C) No expansion

New construction Extension No expansion (Strong effects) (Little effects) (No additional effects)

+40% risk +20% risk Current risk (1 in 650,000 people) (1 in 750,000 people) (1 in 900,000 people)

-30% risk -60% risk Current risk (1 in 4M people) (1 in 7M people) (1 in 3M people)

+200 CHF/year +300 CHF/year +0 CHF/year

2. How certain are you about your choice?

Not certain at all • Not certain • Neither certain nor uncertain • Certain • Very certain • A.14. Decision 2 149

A.14 Decision 2

1. Which option would you prefer?

A) Expansion B) Expansion C) No expansion

Extension New construction No expansion (Little effects) (Strong effects) (No additional effects)

+20% risk +40% risk Current risk (1 in 750,000 people) (1 in 650,000 people) (1 in 900,000 people)

-30% risk -60% risk Current risk (1 in 4M people) (1 in 7M people) (1 in 3M people)

+600 CHF/year +500 CHF/year +0 CHF/year

2. How certain are you about your choice?

Not certain at all • Not certain • Neither certain nor uncertain • Certain • Very certain • 150 Appendix A. Choice Experiment Survey

A.15 Decision 3

1. Which option would you prefer?

A) Expansion B) Expansion C) No expansion

New construction Extension No expansion (Strong effects) (Little effects) (No additional effects)

+40% risk +20% risk Current risk (1 in 650,000 people) (1 in 750,000 people) (1 in 900,000 people)

-60% risk -30% risk Current risk (1 in 7M people) (1 in 4M people) (1 in 3M people)

+100 CHF/year +300 CHF/year +0 CHF/year

2. How certain are you about your choice?

Not certain at all • Not certain • Neither certain nor uncertain • Certain • Very certain • A.16. Decision 4 151

A.16 Decision 4

1. Which option would you prefer?

A) Expansion B) Expansion C) No expansion

New construction Extension No expansion (Strong effects) (Little effects) (No additional effects)

+40% risk +20% risk Current risk (1 in 650,000 people) (1 in 750,000 people) (1 in 900,000 people)

-30% risk -60% risk Current risk (1 in 4M people) (1 in 7M people) (1 in 3M people)

+400 CHF/year +300 CHF/year +0 CHF/year

2. How certain are you about your choice?

Not certain at all • Not certain • Neither certain nor uncertain • Certain • Very certain • 152 Appendix A. Choice Experiment Survey

A.17 Decision 5

1. Which option would you prefer?

A) Expansion B) Expansion C) No expansion

Extension New construction No expansion (Little effects) (Strong effects) (No additional effects)

+40% risk +20% risk Current risk (1 in 650,000 people) (1 in 750,000 people) (1 in 900,000 people)

-30% risk -60% risk Current risk (1 in 4M people) (1 in 7M people) (1 in 3M people)

+500 CHF/year +600 CHF/year +0 CHF/year

2. How certain are you about your choice?

Not certain at all • Not certain • Neither certain nor uncertain • Certain • Very certain • A.18. Decision 6 153

A.18 Decision 6

1. Which option would you prefer?

A) Expansion B) Expansion C) No expansion

Extension New construction No expansion (Little effects) (Strong effects) (No additional effects)

+20% risk +40% risk Current risk (1 in 750,000 people) (1 in 650,000 people) (1 in 900,000 people)

-60% risk -30% risk Current risk (1 in 7M people) (1 in 4M people) (1 in 3M people)

+400 CHF/year +200 CHF/year +0 CHF/year

2. How certain are you about your choice?

Not certain at all • Not certain • Neither certain nor uncertain • Certain • Very certain • 154 Appendix A. Choice Experiment Survey

A.19 Decision 7

1. Which option would you prefer?

A) Expansion B) Expansion C) No expansion

New construction Extension No expansion (Strong effects) (Little effects) (No additional effects)

+20% risk +40% risk Current risk (1 in 750,000 people) (1 in 650,000 people) (1 in 900,000 people)

-60% risk -30% risk Current risk (1 in 7M people) (1 in 4M people) (1 in 3M people)

+300 CHF/year +200 CHF/year +0 CHF/year

2. How certain are you about your choice?

Not certain at all • Not certain • Neither certain nor uncertain • Certain • Very certain • A.20. Decision 8 155

A.20 Decision 8

1. Which option would you prefer?

A) Expansion B) Expansion C) No expansion

New construction Extension No expansion (Strong effects) (Little effects) (No additional effects)

+40% risk +20% risk Current risk (1 in 650,000 people) (1 in 750,000 people) (1 in 900,000 people)

-30% risk -60% risk Current risk (1 in 4M people) (1 in 7M people) (1 in 3M people)

+200 CHF/year +300 CHF/year +0 CHF/year

2. How certain are you about your choice?

Not certain at all • Not certain • Neither certain nor uncertain • Certain • Very certain • 156 Appendix A. Choice Experiment Survey

A.21 Background of your decisions

1. Why have you chosen the "no expansion" option in all choices? [if choice of option C) in all choice tasks]

The proposed changes are too small to be of importance. • I am against an expansion of hydropower. • I don’t think energy policy or hydropower expansion is very impor- • tant. My household cannot afford to pay a higher electricity bill. • I prefer spending my money on other things. • It’s not my household who should pay for this, but others, namely: • ... Other reason: . . . • 2. If you think back about your choices, to what extent did you pay attention to the following characteristics during your decision-making? (0=never considered, 10=always considered)

Type of hydropower expansion • 0 10

Risk of death by a dam breach • 0 10

Risk of death by a nuclear accident • 0 10

Increase in yearly household electricity bill • 0 10

3. If you think back about the possible changes in the risk of dying from a dam breach and a nuclear accident, have you mentally compared these changes with other risks?

Yes • A.22. Your leisure time 157

No • I don’t know • 4. With which other risks have you mentally compared the risk of dying from a dam breach and a nuclear accident? (several answers are possible) [if answer to the previous question is ’yes’ ]

Risk of dying from cancer • Risk of dying from a lightning strike • Other risk: . . . •

A.22 Your leisure time

1. Do you recreate in or along rivers or lakes (for example by ﬁshing, swim- ming and walking)?

Yes, sometimes. About how often do you recreate in or along rivers • or lakes per year? [. . . time(s) per year] No, never • 2. Do hydropower stations negatively affect your enjoyment of recreational activities?

Yes • No • 3. About how many kilometers is it from your place of residence to the next hydropower plant?

Less than 1km • Between 1km and 2km • Between 2km and 5km • Between 5km and 10km • More than 10km • Don’t know • 4. About how many kilometers is it from your place of residence to the next nuclear plant? 158 Appendix A. Choice Experiment Survey

Less than 10km • Between 10km and 20km • Between 20km and 50km • Between 50km and 100km • More than 100km • Don’t know •

A.23 Trust and attitude

1. How much do you trust the following organizations to not just follow their own interests but also represent the interests of the people when it comes to approval procedures for the expansion of hydropower? (0=not at all, 10=very much)

Cantonal authorities • 0 10

Hydropower companies • 0 10

Organizations with primarily conservation interests (nature, environ- • ment and heritage protection)

0 10

Organizations with primarily usage interests (ﬁshing, sport, tourism) • 0 10

2. As you may remember, in 2011 a nuclear accident happened in Fukushima, Japan, involving the release of radiation. Has your attitude towards nuclear energy changed since this event?

Yes • No • 3. How has your attitude changed? (several answers are possible) [if answer to the previous question is ’yes’] A.24. About your person (I/II) 159

I’m against the use of nuclear power since • I since prefer new renewable energy resources like wind and sun • I since prefer hydropower • I since prefer another energy source: . . . • Other reasons: . . . •

A.24 About your person (I/II)

1. Please indicate your gender

Female • Male • 2. In which year were you born?

3. Since which year do you live in Switzerland?

4. Where do you live? Your answer is treated completely anonymously and conﬁdentially and will not be forwarded to third parties.

Street (without the number): . . . • Postal code: . . . • 5. Do you live or have you lived so far in the canton of Aargau?

Yes • No • 6. Where and from when to when? [if answer to the previous question is ’yes’]

From (yyyy): . . . • Till (yyyy): . . . • Postal code: . . . • 7. How many people live in your household? (including yourself and all children) 160 Appendix A. Choice Experiment Survey

A.25 About your person (II/II)

1. What is your highest degree of education?

Mandatory schooling • Diploma (Fachmittelschule, Berufslehre, Berufsmittelschule) • Matura (Gymnasium) • Bachelor degree (University, University of applied sciences, School of • education) Master degree (University, University of applied sciences, School of • education) PhD degree • Other. Which? . . . • 2. What is your current employment status?

Self-employed • Full-time employed • Part-time employed • Retired • Student • Housewife/house husband • Not working • Looking for a job • Other. Which? . . . • 3. Do you consider yourself in general a person who takes risks or do you prefer to avoid taking risks? (0=not at all willing to take risks, 10=very willing to take risks)

0 10

4. Do you currently smoke (at least one cigarette per day)?

Yes • No • A.25. About your person (II/II) 161

5. Do you own shares?

Yes • No • 6. Please indicate in which category your household’s yearly gross income falls. Note: Your gross income is the total income that you and everyone in your household receives in a year (wages and capital income, before taxes and deductions are subtracted). Your answer is treated completely anonymously and conﬁdentially and will not be forwarded to third parties.

Less than 18,000 CHF • 18,001-36,000 CHF • 36,001-54,000 CHF • 54,001-72,000 CHF • 72,001-90,000 CHF • 90,001-108,000 CHF • 108,001-126,000 CHF • 126,001-144,000 CHF • 144,001-162,000 CHF • 162,001-180,000 CHF • 180,001-198,000 CHF • 198,001-216,000 CHF • 216,001-234,000 CHF • 234,001-252,000 CHF • More than 252,000 CHF • 7. Are you the owner of the ﬂat/house you are living in?

Yes • no • 8. Do you or anyone in your household support an environmental organization through membership fees or donations? 162 Appendix A. Choice Experiment Survey

Yes • No •

A.26 On this survey

1. How understandable were the questions in this survey?

Not understandable • Rather not understandable • Partly understandable • Rather understandable • Very understandable • 2. How difﬁcult was it for you to answer the questions in this survey?

Not difficult • Rather not difficult • Partly difficult • Rather difficult • Very difficult • 3. How credible was the information provided in this survey in your opinion?

Not credible • Rather not credible • Partly credible • Rather credible • Very credible • 4. Do you have any other comments concerning this survey? . . .

Thank you very much for answering this survey! For questions and comments please contact [email protected]. 163

Bibliography

Abeler, Johannes et al. (2011). “Reference points and effort provision”. American Economic Review 101 (2), pp. 470–492. Alberini, Anna and Milan Šˇcasný (2011). “Context and the VSL: Evidence from a Stated Preference Study in Italy and the Czech Republic”. Environmental and Resource Economics 49 (4), pp. 511–538. — (2013). “Exploring heterogeneity in the value of a statistical life: Cause of death v. risk perceptions”. Ecological Economics 94, pp. 143–155. Alemu, Mohammed Hussen et al. (2013). “Attending to the Reasons for Attribute Non-attendance in Choice Experiments”. Environmental and Resource Economics 54 (3), pp. 333–359. Alvarez-Farizo, Begona and Nick Hanley (2002). “Using conjoint analysis to quantify public preferences over the environmental impacts of wind farms. An example from Spain”. Energy Policy 30, pp. 107–116. Annaert, Jan et al. (2008). Disposition bias and overconﬁdence in institutional trades. URL: http://www.fma.org/Prague/Papers/Paper_disposition_ bias_and_overconfidence_in_institutional_trades.pdf. Aravena, Claudia, W. George Hutchinson, and Alberto Longo (2012). “Environ- mental pricing of externalities from different sources of electricity generation in Chile”. Energy Economics 34 (4), pp. 1214–1225. Arkes, Hal R. et al. (2008). “Reference point adaptation: Tests in the domain of security trading”. Organizational Behavior and Human Decision Processes 105 (1), pp. 67–81. Arrow, Kenneth et al. (1993). Report of the NOAA Panel on Contingent Valuation. URL: https : / / sites . google . com / site / economiayambiente / PanelNOAA.pdf. Arrow, Kenneth J. and Gerard Debreu (1954). “Existence of an Equilibrium for a Competitive Economy”. Econometrica 22 (3), pp. 265–290. 164 BIBLIOGRAPHY

Baker, Malcolm, Xin Pan, and Jeffrey Wurgler (2010). “A Reference Point Theory of Mergers and Acquisitions”. AFA 2010 Atlanta Meetings Paper. URL: http: //www.nber.org/papers/w15551.pdf. Balcombe, Kelvin, Michael Burton, and Dan Rigby (2011). “Skew and attribute non-attendance within the Bayesian mixed logit model”. Journal of Environ- mental Economics and Management 62 (3), pp. 446–461. Balcombe, Kelvin, Iain Fraser, and Eugene McSorley (2015). “Visual Attention and Attribute Attendance in Multi-Attribute Choice Experiments”. Journal of Applied Econometrics 30 (3), pp. 447–467. Barros, Nathan et al. (2011). “Carbon emission from hydroelectric reservoirs linked to reservoir age and latitude”. Nature Geoscience 4, pp. 593–596. Barry, Michael et al. (2015). “The Future of Swiss Hydropower. A Review on Drivers and Uncertainties”. URL: https://fonew.unibas.ch/fileadmin/ fonew/redaktion/Paper/FoNEW_2015_01.pdf. Bartol, Kathryn M. and David C. Martin (1998). “Applicant Referent Information at Hiring Interview and Subsequent Turnover among Part-Time Workers”. Journal of Vocational Behavior 53 (3), pp. 334–352. Bateman, Ian J. and Roy Brouwer (2006). “Consistency and construction in stated WTP for health risk reductions: A novel scope-sensitivity test”. Resource and Energy Economics 28 (3), pp. 199–214. Bateman, Ian J., Alistair Munro, and Gregory L. Poe (2008). “Decoy effects in choice experiments and contingent valuation: asymmetric dominance”. Land Economics 84 (1), pp. 115–127. Bateman, Ian J. et al. (2002). Economic Valuation with Stated Preference Techniques. Cheltenham UK: Edward Elgar. Beck, Matthew J., Simon Fifer, and John M. Rose (2016). “Can you ever be certain? Reducing hypothetical bias in stated choice experiments via respondent reported choice certainty”. Transportation Research Part B: Methodological 89, pp. 149–167. Beck, Matthew J., John M. Rose, and David A. Hensher (2013). “Consistently inconsistent: The role of certainty, acceptability and scale in choice”. Trans- portation Research Part E: Logistics and Transportation Review 56, pp. 81–93. Bergland, Olvar (1998). Valuing Aesthetical Values of Weirs in Watercourses with Hydroelectric Plants - Verdsetjing av estetiske verdiar i tilknytning til tersklar i regulerte vassdrag. Oslo: Norwegian Water Resources and Energy Directorate (NVE). BIBLIOGRAPHY 165

Bergmann, Ariel, Sergio Colombo, and Nick Hanley (2008). “Rural versus urban preferences for renewable energy developments”. Ecological Economics 65, pp. 616–625. Bergmann, Ariel, Nick Hanley, and Robert Wright (2006). “Valuing the attributes of renewable energy investments”. Energy Policy 34 (9), pp. 1004–1014. Bier, V.M (2001). “On the state of the art: risk communication to the public”. Reliability Engineering & System Safety 71 (2), pp. 139–150. Biro, Yasemin E. K. (1998). “Valuation of the Environmental Impacts of the Kayrak- tepe Dam/Hydroelectric Project, Turkey: An Exercise in Contingent Valua- tion”. Ambio 27 (3), pp. 224–229. Bliem, Markus, Michael Getzner, and Petra Rodiga-Laßnig (2012). “Temporal stability of individual preferences for river restoration in Austria using a choice experiment”. Journal of Environmental Management 103, pp. 65–73. Bogliacino, Francesco and Iván González-Gallo (2015). “Aspirations, Prospect Theory and entrepreneurship: evidence from Colombia”. International Review of Economics 62 (3), pp. 271–290. Born, Stephen M. et al. (1998). “Socioeconomic and Institutional Dimensions of Dam Removals: The Wisconsin Experience”. Environmental Management 22 (3), pp. 359–370. Bothe, David (2003). Environmental Costs due to the Karahnjukar Hydro Power Project on Iceland. University of Cologne: Department of Economic and Social Geog- raphy, Cologne, Germany. Botzen, W. J. W. and J. C. J. M. van den Bergh (2012). “Risk attitudes to low- probability climate change risks: WTP for ﬂood insurance”. Journal of Eco- nomic Behavior and Organization 82 (1), pp. 151–166. Box, G. E. P. and D. R. Cox (1964). “An Analysis of Transformations”. Journal of the Royal Statistical Society. Series B (Methodological) 26 (2), pp. 211–252. Boyd, James and Spencer Banzhaf (2007). “What are ecosystem services? The need for standardized environmental accounting units”. Ecological Economics 63 (2-3), pp. 616–626. Brander, Luke, Roy Brouwer, and Alfred Wagtendonk (2013). “Economic valuation of regulating services provided by wetlands in agricultural landscapes: A meta-analysis”. Ecological Engineering 56, pp. 89–96. Brouwer, Roy (2000). “Environmental Value Transfer: State of the Art and Future Prospects”. Ecological Economics 32 (1), pp. 137–152. 166 BIBLIOGRAPHY

Brouwer, Roy, Ivana Logar, and Oleg Sheremet (2016). “Choice Consistency and Preference Stability in Test-Retests of Discrete Choice Experiment and Open- Ended Willingness to Pay Elicitation Formats”. Environmental and Resource Economics, pp. 1–23. Brouwer, Roy and Julia Martín-Ortega (2012). “Modeling self-censoring of pol- luter pays protest votes in stated preference research to support resource damage estimations in environmental liability”. Resource and Energy Economics 34 (1), pp. 151–166. Brouwer, Roy et al. (2010). “Choice Certainty and Consistency in Repeated Choice Experiments”. Environmental and Resource Economics 46, pp. 93–109. Brown, Thomas C. et al. (2008). “Reliability of individual valuations of public and private goods: Choice consistency, response time, and preference reﬁne- ment”. Journal of Public Economics 92 (7), pp. 1595–1606. Bruce, Amanda, John Crespi, and Jayson Lusk (2015). “The behavioral and neuroeconomics of food and brand decisions: Executive summary”. Journal of Agricultural and Food Industrial Organization 13 (1), pp. 1–4. Bruno, Maria Cristina et al. (2009). “Impact of hydropeaking on hyporheic invertebrates in an Alpine stream (Trentino, Italy)”. International Journal of Limnol- ogy 45 (3), pp. 157–170. Burgherr, Peter and Stefan Hirschberg (2008). “A Comparative Analysis of Acci- dent Risks in Fossil, Hydro, and Nuclear Energy Chains”. Human and Ecolog- ical Risk Assessment: An International Journal 14 (5), pp. 947–973. — (2014). “Comparative risk assessment of severe accidents in the energy sec- tor”. Energy Policy 74, S45–S56. Camerer, Colin et al. (1997). “Labor Suply of New York City Cabdrivers: One Day at a Time”. The Quarterly Journal of Economics 112 (2), pp. 407–441. Cameron, A. Colin and Pravin K. Trivedi (2005). Microeconometrics: Methods and Applications. New York, NY: Cambridge University Press. Campbell, Danny (2007). “Willingness to pay for rural landscape improvements: Combining mixed logit and random-effects model”. Journal of Agricultural Economics 58 (3), pp. 467–483. Campbell, Danny, David A. Hensher, and Riccardo Scarpa (2011). “Non-attendance to attributes in environmental choice analysis: a latent class speciﬁcation”. Journal of Environmental Planning and Management 54 (8), pp. 1061–1076. Campbell, Danny, W. George Hutchinson, and Riccardo Scarpa (2008). “Incor- porating discontinuous preferences into the analysis of discrete choice experiments”. Environmental and Resource Economics 41 (3), pp. 401–417. BIBLIOGRAPHY 167

Campbell, Danny, Morten Raun Mørkbak, and Søren Bøye Olsen (2017). “Re- sponse time in online stated choice experiments: the non-triviality of identifying fast and slow respondents”. Journal of Environmental Economics and Pol- icy 6 (1), pp. 17–35. URL: http://dx.doi.org/10.1080/21606544. 2016.1167632. Carlsson, Fredrik, Mitesh Kataria, and Elina Lampi (2010). “Dealing with ignored attributes in choice experiments on valuation of Sweden’s environmental quality objectives”. Environmental and Resource Economics 47 (1), pp. 65–89. Carlsson, Fredrik, Morten Raun Mørkbak, and Søren Bøye Olsen (2012). “The ﬁrst time is the hardest: A test of ordering effects in choice experiments”. Journal of Choice Modelling 5 (2), pp. 19–37. Carson, Richard T. (1997). “Contingent valuation surveys and tests of insensitivity to scope”. Determining the Value of Non-Marketed Goods: Economic, Psycho- logical, and Policy Relevant Aspects of Contingent Valuation Methods. Ed. by R.J. Kopp, W.W. Pommerehne, and N. Schwarz. Boston, MA: Kluwer Academic Publishers. Carson, Richard T. and Robert C. Mitchell (1993). “The Issue of Scope in Con- tingent Valuation Studies”. American Journal of Agricultural Economics 75 (5), pp. 1263–1267. Carter, Steven and Michael McBride (2013). “Experienced utility versus decision utility: Putting the ’S’ in satisfaction”. Journal of Socio-Economics 42, pp. 13–23. Chapman, Gretchen B. (2000). “Preferences for improving and declining sequences of health outcomes”. Journal of Behavioral Decision Making 13, pp. 203–218. Charness, Gary and Uri Gneezy (2012). “Strong Evidence for Gender Differences in Risk Taking”. Journal of Economic Behavior and Organization 83 (1), pp. 50–58. Chen, Haipeng (Allan) and Akshay R. Rao (2002). “Close Encounters of Two Kinds: False Alarms and Dashed Hopes”. Marketing Science 21, pp. 178–196. ChoiceMetrics (2012). Ngene Manual 1.1.2. User Manual & Reference Guide. URL: http://www.choice-metrics.com/download.html. Collins, Andrew T., John M. Rose, and David A. Hensher (2013). “Speciﬁcation issues in a generalised random parameters attribute nonattendance model”. Transportation Research Part B: Methodological 56, pp. 234–253. Colombo, Sergio, Michael Christie, and Nick Hanley (2013). “What are the consequences of ignoring attributes in choice experiments? Implications for ecosystem service valuation”. Ecological Economics 96, pp. 25–35. 168 BIBLIOGRAPHY

Colombo, Sergio, Klaus Glenk, and Beatriz Rocamora-Montiel (2016). “Analysis of choice inconsistencies in on-line choice experiments: impact on welfare measures”. European Review of Agricultural Economics 43 (2), pp. 271–302. Connelly, Nancy A. and Barbara A. Knuth (1998). “Evaluating risk communication: examining target audience perceptions about four presentation formats for ﬁsh consumption health advisory information.” Risk analysis 18 (5), pp. 649–659. Corso, Phaedra S., James K. Hammitt, and John D. Graham (2001). “Valuing Mortality-Risk Reduction: Using Visual Aids to Improve the Validity of Con- tingent Valuation”. Journal of Risk and Uncertainty 23 (2), pp. 165–184. Crawford, Vincent P. and Juanjuan Meng (2011). “New York City Cabdrivers’ Labor Supply Revisited: Reference-Dependence Preferences with Rational- Expectations Targets for Hours and Income”. American Economic Review 101, pp. 1912–1932. De Bekker-Grob, Esther W., Mandy Ryan, and Karen Gerard (2012). “Discrete Choice Experiments in Health Economics: A Review of the Literature”. Health economics 21, pp. 145–172. De Borger, Bruno and Mogens Fosgerau (2008). “The trade-off between money and travel time: A test of the theory of reference-dependent preferences”. Journal of Urban Economics 64 (1), pp. 101–115. De Moraes Ramos, Giselle, Winnie Daamen, and Serge Hoogendoorn (2013). “Modelling travellers’ heterogeneous route choice behaviour as prospect max- imizers”. Journal of Choice Modelling 6, pp. 17–33. Debreu, Gerard (1959). Theory of Value. New York: John Wiley. Dekker, Thijs, Paul Koster, and Roy Brouwer (2014). “Changing with the tide: Semi-parametric estimation of preference dynamics”. Land Economics 90 (4), pp. 717–745. Dekker, Thijs et al. (2011). “The Effect of Risk Context on the Value of a Statistical Life: A Bayesian Meta-model”. Environmental and Resource Economics 49 (4), pp. 597–624. Dekker, Thijs et al. (2016). “Decision uncertainty in multi-attribute stated preference studies”. Resource and Energy Economics 43, pp. 57–73. Delsontro, Tonya et al. (2010). “Extreme Methane Emissions from a Swiss Hy- dropower Reservoir: Contribution from Bubbling Sediments”. Environmental Science & Technology 44, pp. 2419–2425. BIBLIOGRAPHY 169

Desvousges, William, Kristy Mathews, and Kenneth Train (2012). “Adequate re- sponsiveness to scope in contingent valuation”. Ecological Economics 84, pp. 121– 128. Determann, Domino et al. (2017). “Impact of Survey Administration Mode on the Results of a Health-Related Discrete Choice Experiment: Online and Paper Comparison”. Value in Health (in press). Devine-Wright, Patrick (2005). “Beyond NIMBYism: towards an Integrated Frame- work for Understanding Public Perceptions of Wind Energy”. Wind Energy 8, pp. 125–139. Dillard, James Price (1994). “Rethinking the Study of Fear Appeals: An Emo- tional Perspective”. Communication Theory 4, pp. 295–323. Efron, Bradley and Gail Gong (1983). “A leisurely look at the bootstrap, the jack- knife, and cross-validation”. The American Statistician 37 (1), pp. 36–48. Ehrlich, Üllas and Mart Reimann (2010). “Hydropower versus Non-market Val- ues of Nature : A Contingent Valuation Study of Jägala Waterfalls , Estonia”. International Journal of Geology 4 (3), pp. 59–63. Ek, Kristina (2006). “Quantifying the environmental impacts of renewable energy: the case of Swedish wind power”. Environmental Valuation in Develped Countries: Case Studies. Ed. by David Pearce. Cheltenham, UK: Edward Elgar Publishing Limited, pp. 181–212. Ek, Kristina and Lars Persson (2014). “Wind farms – Where and how to place them? A choice experiment approach to measure consumer preferences for characteristics of wind farm establishments in Sweden”. Ecological Economics 105, pp. 193–203. ElCom (2014). Strompreise 2015: Durchschnittlich steigende Tarife für Haushalte und mittlere Betriebe. Bern. URL: https://www.admin.ch/gov/de/start/ dokumentation/medienmitteilungen.msg-id-54336.html. Fafchamps, Marcel, Bereket Kebede, and Daniel John Zizzo (2015). “Keep up with the winners: Experimental evidence on risk taking, asset integration, and peer effects”. European Economic Review 79, pp. 59–79. Faiers, Adam and Charles Neame (2006). “Consumer attitudes towards domestic solar power systems”. Energy Policy 34 (14), pp. 1797–1806. Fehr, Ernst and Lorenz Goette (2007). “Work More if Wages Are High ? Evidence from a Randomized Field Experiment”. The American Economic Review 97 (1), pp. 298–317. Fiegenbaum, Avi (1990). “Prospect theory and the risk-return association”. Jour- nal of Economic Behavior and Organization 14 (2), pp. 187–203. 170 BIBLIOGRAPHY

Filippini, Massimo, Loa Buchli, and Silvia Banfi (2003). “Estimating the benefits of low flow alleviation in rivers: the case of the Ticino River”. Applied Eco- nomics 35, pp. 585–590. Finucane, Melissa L. et al. (2000). “The affect heuristic in judgments of risks and benefits”. Journal of Behavioral Decision Making 13 (1), pp. 1–17. Foster, V. and Susana Mourato (2002). “Testing for consistency in contingent ranking experiments”. Journal of Environmental Economics and Management 44, pp. 309–328. Gau, R. and M. Viswanathan (2008). “The Retail Shopping Experience for Low- Literate Consumers”. Journal of Research for Consumers 15, pp. 1–8. Genesove, David and Christopher Mayer (2001). “Loss Aversion and Seller Be- havior: Evidence from the Housing Market”. Quarterly Journal of Economics 116, pp. 1233–1260. Getzner, Michael (2015). “Importance of Free-Flowing Rivers for Recreation: Case Study of the River Mur in Styria, Austria”. Journal of Water Resources Planning and Management 141 (2), p. 04014050. Gneezy, Uri (2005). “Updating the Reference Level: Experimental Evidence”. Ex- perimental Business Reserach. Ed. by R. Zwick and A. Rapoport. Netherlands: Springer, pp. 263–284. Gogniat, Sandra (2011). Estimating the Benefits of an Improvement in Water Qual- ity and Flow Regulation : Case study of the Doubs. Master’s Thesis, Université de Neuchâtel, Neuchâtel, Switzerland. URL: https://www.unine.ch/ files/live/sites/iaf/files/shared/documents/Seminairesenfinance/ Seminaireseconomieetfinance/SG_Thesis_final.pdf. Gómez-Mejía, Luis R. et al. (2007). “Socioemotional Wealth and Business Risks in Family-controlled Firms: Evidence from Spanish Olive Oil Mills”. Admin- istrative Science Quarterly 52, pp. 106–137. Greenberg, Michael and Heather Barnes Truelove (2011). “Energy choices and risk beliefs: Is it just global warming and fear of a nuclear power plant accident?” Risk Analysis 31 (5), pp. 819–831. Greene, William H (2012). LIMDEP Version 10 Econometric Modeling Guide. URL: http://people.stern.nyu.edu/wgreene/DiscreteChoice/Software/ LIMDEP10ModelingGuide.pdf. Guo, Xiurui et al. (2014). “Willingness to pay for renewable electricity: A contingent valuation study in Beijing, China”. Energy Policy 68, pp. 340–347. BIBLIOGRAPHY 171

Hack, Andreas and Frauke von Bieberstein (2014). “How expectations affect reference point formation: an experimental investigation”. Review of Managerial Science 9, pp. 33–59. Hack, Andreas, Frauke von Bieberstein, and Nils D. Kraiczy (2015). “Reference point formation and new venture creation”. Small Business Economics 46 (3), pp. 447–465. Håkansson, Cecilia (2009). “Costs and beneﬁts of improving wild salmon passage in a regulated river”. Journal of Environmental Planning and Management 52 (3), pp. 345–363. Hamilton, L. C. (2012). Statistics with STATA Version 12. Boston, MA: Cengage Learning. Han, Sang-Yong, Seung-Jun Kwak, and Seung-Hoon Yoo (2008). “Valuing environmental impacts of large dam construction in Korea: An application of choice experiments”. Environmental Impact Assessment Review 28 (4-5), pp. 256– 266. Hanley, Nick and Ceara Nevin (1999). “Appraising renewable energy developments in remote communities: the case of the North Assynt Estate, Scotland”. Energy Policy 27 (9), pp. 527–547. Hanley, Nick, Robert E. Wright, and Vic Adamowicz (1998). “Using Choice Ex- periments to Value the Environment”. Design Issues, Current Experience and Future Prospects 11 (3-4), pp. 413–428. Hanley, Nick, Robert E. Wright, and Gary Koop (2002). “Modelling recreation demand using choice experiments: Climbing in Scotland”. Environmental and Resource Economics 22 (3), pp. 449–466. Hansesveen, H. and G. Helgas (1997). Environmental Costs of Hydropower Devel- opment – Estimering av miljokostnader ved en vannkraftutbygging i Ovre Otta. Norwegian University of Life Sciences, As, Norway. Hansson, Helena and Carl Johan Lagerkvist (2014). “Decision Making for Ani- mal Health and Welfare: Integrating Risk-Beneﬁt Analysis with Prospect The- ory”. Risk Analysis 34 (6), pp. 1149–1159. Harbord, Roger M. and Julian P. T. Higgins (2008). “Meta-regression in Stata”. The Stata Journal 8 (4), pp. 493–519. Harrison, Glenn W., Morten I. Lau, and B. Melonie (2000). “Estimating Indi- vidual Discount Rates in Denmark: A Field Experiment”. Business 92 (5), pp. 1606–1617. 172 BIBLIOGRAPHY

Hartmann, Patrick et al. (2013). “Nuclear power threats, public opposition and green electricity adoption: Effects of threat belief appraisal and fear arousal”. Energy Policy 62, pp. 1366–1376. He, Xue Dong and Xun Yu XY Zhou (2014). “Myopic loss aversion, reference point, and money illusion”. Quantitative Finance 14 (9), pp. 1541–1554. Heath, Chip, Steven Huddart, and Mark Lang (1999). “Psychological Factors and Stock Option Exercise”. Quarterly Journal of Economics 114, pp. 601–627. Heath, Chip, Richard P Larrick, and George Wu (1999). “Goals as Reference Points”. Cognitive Psychology 38, pp. 79–109. Helena, Annina et al. (2015). “Renewable Energy, Authenticity, and Tourism: Social Acceptance of Photovoltaic Installations in a Swiss Alpine Region”. Mountain Research and Development 35 (2), pp. 161–170. Hensher, D. A. and William H. Greene (2002). “The Mixed Logit Model: The State of Practice”. Transportation 30, pp. 133–176. Hensher, David A., Andrew T. Collins, and William H. Greene (2013). “Account- ing for attribute non-attendance and common-metric aggregation in a probabilistic decision process mixed multinomial logit model: A warning on potential confounding”. Transportation 40 (5), pp. 1003–1020. Hensher, David A. and William H. Greene (2010). “Non-attendance and dual processing of common-metric attributes in choice analysis: a latent class specification”. Empirical Economics 39 (2), pp. 413–426. Hensher, David A. and John M. Rose (2009). “Simplifying choice through attribute preservation or non-attendance: Implications for willingness to pay”. Transportation Research Part E: Logistics and Transportation Review 45 (4), pp. 583– 590. Hensher, David A., John M. Rose, and Matthew J. Beck (2012). “Are there specific design elements of choice experiments and types of people that influence choice response certainty?” Journal of Choice Modelling 5 (1), pp. 77–97. Hensher, David A., John M. Rose, and Tony Bertoia (2007). “The implications on willingness to pay of a stochastic treatment of attribute processing in stated choice studies”. Transportation Research Part E: Logistics and Transportation Re- view 43 (2), pp. 73–89. Hensher, David A., John M. Rose, and William H. Greene (2005). “The implications on willingness to pay of respondents ignoring specific attributes”. Transportation 32 (3), pp. 203–222. BIBLIOGRAPHY 173

— (2012). “Inferring attribute non-attendance from stated choice data: Implica- tions for willingness to pay estimates and a warning for stated choice experiment design”. Transportation 39 (2), pp. 235–245. — (2015). Applied Choice Analysis. Cambridge UK: Cambridge University Press. Hess, Stephane and David A. Hensher (2010). “Using conditioning on observed choices to retrieve individual-speciﬁc attribute processing strategies”. Trans- portation Research Part B: Methodological 44 (6), pp. 781–790. — (2013). “Making use of respondent reported processing information to understand attribute importance: A latent variable scaling approach”. Transporta- tion 40 (2), pp. 397–412. Hess, Stephane, John M. Rose, and David A. Hensher (2008). “Asymmetric preference formation in willingness to pay estimates in discrete choice models”. Transportation Research Part E: Logistics and Transportation Review 44, pp. 847– 863. Hess, Stephane et al. (2013). “It’s not that I don’t care, I just don’t care very much: Confounding between attribute non-attendance and taste heterogeneity”. Transportation 40 (3), pp. 583–607. Heyman, James et al. (2004). “I was pleased a moment ago: How pleasure varies with background and foreground reference points”. Motivation and Emotion 28 (1), pp. 65–83. Hirschberg, Stefan et al. (2016). “Health effects of technologies for power generation: Contributions from normal operation, severe accidents and terrorist threat”. Reliability Engineering & System Safety 145, pp. 373–387. Hökby, Stina and Tore Söderqvist (2003). “Elasticities of Demand and Willing- ness to Pay for Environmental Services in Sweden”. Environmental and Re- source Economics 26 (3), pp. 361–383. Hole, Arne Risa (2011). “A discrete choice model with endogenous attribute attendance”. Economics Letters 110 (3), pp. 203–205. Hole, Arne Risa, Julie Riise Kolstad, and Dorte Gyrd-Hansen (2013). “Inferred vs. stated attribute non-attendance in choice experiments: A study of doc- tors’ prescription behaviour”. Journal of Economic Behavior and Organization 96, pp. 21–31. Hynes, Stephen and Nick Hanley (2006). “Preservation versus development on Irish rivers: whitewater kayaking and hydro-power in Ireland”. Land Use Pol- icy 23 (2), pp. 170–180. 174 BIBLIOGRAPHY

IAEA (International Atomic Energy Agency) (2015). Nuclear Power Reactors in the World. URL: http://www-pub.iaea.org/MTCD/Publications/PDF/ rds2-35web-85937611.pdf. IEA (International Energy Agency) (2010). Renewable Energy Essentials: Hydropower. URL: http : / / www . iea . org / publications / freepublications / publication/Hydropower_Essentials.pdf. — (2014a). Energy Statistics of Non-OECD Countries. Paris. URL: http://www. oecd-ilibrary.org/energy/energy-statistics-of-non-oecd- countries-2014_energy_non-oecd-2014-en. — (2014b). Energy Statistics of OECD Countries. Paris. URL: http://www.oecd- ilibrary.org/energy/energy-balances-of-oecd-countries- 2014_energy_bal_oecd-2014-en. — (2017). IEA finds CO2 emissions flat for third straight year even as global economy grew in 2016. URL: https://www.iea.org/newsroom/news/2017/ march/iea-finds-co2-emissions-flat-for-third-straight- year-even-as-global-economy-grew.html. IMF (International Monetary Fund) (2014). Proposed New Grouping in WEO Coun- try Classifications: Low-Income Developing Countries. IMF Policy Paper. Washing- ton, DC. URL: https://www.imf.org/external/np/pp/eng/2014/ 060314.pdf. — (2017). World Economic Outlook Update January 2017. URL: https://www. imf.org/external/pubs/ft/weo/2017/update/01/. Jacobson, Mark Z. and Mark A. Delucchi (2009). “A Path to Sustainable Energy by 2030”. Scientific American 301, pp. 58–65. Jehle, Geoffrey and Philip Reny (2001). Advanced Microeconomic Theory. Boston: Addison Wesley. Johansson, Maria and Thorbjörn Laike (2007). “Intention to respond to local wind turbines: The role of attitudes and visual perception”. Wind Energy 10 (5), pp. 435–451. Johnston, Robert J. et al. (2017). “Contemporary Guidance for Stated Preference Studies”. Journal of the Association of Environmental and Resource Economists 4 (2), pp. 319–405. Kahneman, Daniel and Amos Tversky (1979). “Prospect Theory: An Analysis of Decision under Risk”. Econometrica 47 (2), pp. 263–292. Kataria, Mitesh (2009). “Willingness to pay for environmental improvements in hydropower regulated rivers”. Energy Economics 31 (1), pp. 69–76. BIBLIOGRAPHY 175

Kehlbacher, Ariane, Kelvin Balcombe, and Richard Bennett (2013). “Stated attribute non-attendance in successive choice experiments”. Journal of Agricul- tural Economics 64 (3), pp. 693–706. Keller, Carmen, Michael Siegrist, and Vivianne Visschers (2009). “Effect of risk ladder format on risk perception in high- and low-numerate individuals”. Risk Analysis 29 (9), pp. 1255–1264. Kirby, Kris N. and R. J. Herrnstein (1995). “Preference Reversals Due To Myopic Discounting of Delayed Reward”. Psychological Science 6 (2), pp. 83–89. Klinglmair, Andrea and Markus Bliem (2013). Die Erschliessung vorhandener Wasserkraft- potenziale in Österreich im Spannungsfeld von Energiepolitik und ökologischen Schutzzie- len. 8. Internationale Energiewirtschaftstagung an der TU Wien. URL: http: //eeg.tuwien.ac.at/eeg.tuwien.ac.at_pages/events/iewt/ iewt2015/uploads/fullpaper/P_146_Klinglmair_Andrea_8- Jan-2013_10:59.pdf. Klinglmair, Andrea, Markus Bliem, and Roy Brouwer (2012). Public Preferences for Urban and Rural Hydropower Projects in Styria using a Choice Experiment. IHS Kärnten Working Paper. URL: http://www.carinthia.ihs.ac.at/ HydroVal/files/working_paper.pdf. Koetse, Mark J. and Roy Brouwer (2015). “Reference Dependence Effects on WTA and WTP Value Functions and Their Disparity”. Environmental and Resource Economics 65, pp. 723–745. Komarek, Timothy M., Frank Lupi, and Michael D. Kaplowitz (2011). “Valuing energy policy attributes for environmental management: Choice experiment evidence from a research institution”. Energy Policy 39 (9), pp. 5105–5115. Kosenius, Anna-Kaisa (2009). “Causes of Response uncertainty and its Implica- tions for WTP Estimation in Choice Experiment”. Discussion Papers No. 29, University of Helsinki. Kosenius, Anna-Kaisa and Markku Ollikainen (2013). “Valuation of environmental and societal trade-offs of renewable energy sources”. Energy Policy 62, pp. 1148–1156. Kosz, M (1996). “Valuing riverside wetlands: the case of the “Donau-Auen” national park”. Ecological Economics 16, pp. 109–127. Koszegi, Botond and Matthew Rabin (2006). “A Model of Reference-Dependent Preferences”. The Quarterly Journal of Economics 121 (4), pp. 1133–1165. Kragt, Marit E. (2013). “Stated and inferred attribute attendance models: A comparison with environmental choice experiments”. Journal of Agricultural Eco- nomics 64 (3), pp. 719–736. 176 BIBLIOGRAPHY

Krinsky, Itzhak and A. Leslie Robb (1986). “On Approximating the Statistical Properties of Elasticities”. The Review of Economics and Statistics 68 (4), pp. 715– 719. Ku, Se-Ju and Seung-Hoon Yoo (2010). “Willingness to pay for renewable energy investment in Korea: A choice experiment study”. Renewable and Sustainable Energy Reviews 14 (8), pp. 2196–2201. Ladenburg, Jacob and Alex Dubgaard (2007). “Willingness to pay for reduced visual disamentities from offshore wind farms in Denmark”. Energy Policy 35, pp. 4059–4071. Ladenburg, Jacob and Søren Bøye Olsen (2008). “Gender-speciﬁc starting point bias in choice experiments: Evidence from an empirical study”. Journal of En- vironmental Economics and Management 56 (3), pp. 275–285. Lagarde, Mylene (2010). “Investigating attribute non-attendance and its consequences in choice experiments with latent class models”. Health economics 22, pp. 554–567. Lahey, Joanna N. and Douglas Oxley (2016). “The Power of Eye Tracking in Eco- nomics Experiments”. American Economic Review 106 (5), pp. 309–313. Lancaster, Kevin J. (1966). “A New Approach to Consumer Theory”. The Journal of Political Economy 74 (2), pp. 132–157. Lancsar, Emily and Jordan Louviere (2006). “Deleting ’irrational’ responses from discrete choice experiments: A case of investigating or imposing preferences?” Health Economics 15 (8), pp. 797–811. Lant, Theresa K. and Stephen J. Mezias (1992). “An Organizational Learning Model of Convergence and Reorientation”. Organization Science 3 (1), pp. 47– 71. Lanz, Bruno et al. (2010). “Investigating Willingness to Pay - Willingness to Ac- cept Asymmetry in Choice Experiments”. Choice modeling: The state-of-the-art and the state-of-practice. Ed. by Stephane Hess and Andrew Daly. Bingley, UK: Emerald Group Publishing Limited. Liebe, Ulf, Jürgen Meyerhoff, and Volkmar Hartje (2012). “Test-Retest Reliabil- ity of Choice Experiments in Environmental Valuation”. Environmental and Resource Economics 53 (3), pp. 389–407. Lienhoop, Nele and Douglas MacMillan (2007). “Valuing wilderness in Iceland: Estimation of WTA and WTP using the market stall approach to contingent valuation”. Land Use Policy 24 (1), pp. 289–295. Lipkus, I. M. and J. G. Hollands (1999). “The visual communication of risk.” Jour- nal of the National Cancer Institute. Monographs 27701 (25), pp. 149–163. BIBLIOGRAPHY 177

Lipsey, Mark W. and David B. Wilson (2001). Practical Meta-Analysis. Thousand Oaks, London, New Delhi: Sage Publications. Logar, Ivana and Roy Brouwer (2017). “The Effect of Risk Communication on Choice Behavior, Welfare Estimates and Choice Certainty”. Water Resources and Economics 18, pp. 34–50. Lohse, Gerald L. and Eric J. Johnson (1996). “A Comparison of Two Process Trac- ing Methods for Choice Tasks”. Organizational Behavior and Human Decision Processes 68 (1), pp. 28–43. Longo, Alberto, Anil Markandya, and Marta Petrucci (2008). “The internalization of externalities in the production of electricity: Willingness to pay for the attributes of a policy for renewable energy”. Ecological Economics 67 (1), pp. 140–152. Loomis, John (1996). “Measuring the economic benefits of removing dams and restoring the Elwha River: Results of a contingent valuation survey”. Water Resources Research 32 (2), pp. 441–447. — (2002). “Quantifying recreation use values from removing dams and restoring free-flowing rivers: A contingent behavior travel cost demand model for the Lower Snake River”. Water Resources Research 38 (6), pp. 2–1. Loomis, John, Cindy Sorg, and Dennis Donnelly (1986). “Economic Losses to Recreational Fisheries due to Small-head Hydro-power Development: a Case Study of the Henry’s Fork in Idaho”. Journal of Environmental Management 22 (1), pp. 85–94. Loomis, John B. and Pierre H. DuVair (1993). “Evaluating the Effect of Alterna- tive Risk Communication Devices on Willingness to Pay: Results from a Di- chotomous Choice Contingent Valuation Experiment”. Land Economics 69 (3), pp. 287–298. Louviere, Jordan J., David A. Hensher, and Joffre D. Swait (2000). Stated choice methods: analysis and application. Cambridge: Cambridge University Press. Lundhede, Thomas Hedemark et al. (2009). “Handling respondent uncertainty in Choice Experiments: Evaluating recoding approaches against explicit modelling of uncertainty”. Journal of Choice Modelling 2 (2), pp. 118–147. Lusk, Jayson L. et al. (2016). “Neural antecedents of a random utility model”. Journal of Economic Behavior & Organization 132, Part, pp. 93–103. Maddux, James E. and Ronald W. Rogers (1983). “Protection motivation and self- efficacy: A revised theory of fear appeals and attitude change”. Journal of Ex- perimental Social Psychology 19 (5), pp. 469–479. 178 BIBLIOGRAPHY

Mas-Colell, Andreu and Michael D. Whinston (1995). Microeconomic Theory. New York: Oxford University Press. Mattmann, Matteo, Ivana Logar, and Roy Brouwer (2016a). “Hydropower externalities: A meta-analysis”. Energy Economics 57, pp. 66–77. — (2016b). “Wind power externalities: A meta-analysis”. Ecological Economics 127, pp. 23–36. McDonald, Rebecca L. et al. (2016). “Dread and latency impacts on a VSL for cancer risk reductions”. Journal of Risk and Uncertainty 52 (2), pp. 137–161. McFadden, Daniel (1974). “Conditional logit analysis of qualitative choice behaviour”. Frontiers in econometrics. Ed. by P. Zarembka. New York: Academic Press. — (1979). “Quantitative Methods for Analyzing Travel Behaviour on Individu- als: Some Recent Developments”. Behavioral Travel Modelling. Ed. by David A. Hensher and Peter R. Stopher. Kent: Croom Helm, pp. 279–318. — (2001). “Economic Choices”. The American Economic Review 91 (3), pp. 351– 378. Medvec, Victoria Husted, Scott F. Madey, and Thomas Gilovich (1995). “When less is more: Counterfactual thinking and satisfaction among olympic medal- ist”. Journal of Personality and Social Psychology 69 (4), pp. 603–610. Meissner, Martin, Sören W. Scholz, and Reinhold Decker (2010). “Using Eye Track- ing and Mouselab to Examine how Respondents process Information in CBC”. Proceedings of the Sawtooth Software Conference, pp. 151–171. Mellers, Barbara, Alan Schwartz, and Ilana Ritov (1999). “Emotion-Based Choice”. Journal of Experimental Psychology: General 128 (3), pp. 332–345. Menegaki, Angeliki N., Søren Bøye Olsen, and Konstantinos P. Tsagarakis (2016). “Towards a common standard – A reporting checklist for web-based stated preference valuation surveys and a critique for mode surveys”. Journal of Choice Modelling 18, pp. 18–50. Meyers-Levy, Joan (1989). “Gender differences in information processing: a selective interpretation”. Cognitive and Affective Responses to Advertising. Ed. by Patricia Cafferata and Alice M. Tybout. Lanham, Maryland: Lexington. Meyers-Levy, Joan and Barbara Loken (2015). “Revisiting gender differences: What we know and what lies ahead”. Journal of Consumer Psychology 25 (1), pp. 129–149. BIBLIOGRAPHY 179

Mørkbak, Morten Raun and Søren Bøye Olsen (2015). “A within-sample investigation of test-retest reliability in choice experiment surveys with real economic incentives”. Australian Journal of Agricultural and Resource Economics 59 (3), pp. 375–392. Navrud, Ståle (1995). Hydro Fuel Cycle. Part II (p.127-249) in European Commission DG XII Science Research and Innovation (1995): ExternE: Externalities of Energy. Volume 6: Wind and Hydro. EUR 16525 EN, European Comission Publishing. Lux- embourg. — (2001). “Environmental costs of hydro compared with other energy options”. Hydropower and Dams 8 (2), pp. 44–48. — (2004). Environmental Costs of Hydropower, Second Stage - Miljøkostnadsprosjektet Trinn 2. EBL report 181. Nelson, Jon P. and Peter E. Kennedy (2008). “The Use (and Abuse) of Meta- Analysis in Environmental and Natural Resource Economics: An Assessment”. Environmental and Resource Economics 42 (3), pp. 345–377. Nguyen, Thanh Cong et al. (2015). “Attribute non-attendance in discrete choice experiments: A case study in a developing country”. Economic Analysis and Policy 47, pp. 22–33. Odean, T. (1998). “Are Investors Reluctant to Realise their Losses?” The Journal of Finance 53 (5), pp. 1775–1798. OECD (2014). OECD. Statistics (database). URL: http://www.oecd-ilibrary. org/statistics. — (2017). Exchange rates (indicator). URL: https://data.oecd.org/conversion/ exchange-rates.htm. Ohdoko, Taro (2008). “Comparison of Complete Combinatorial and Likelihood Ratio Tests: Empirical Findings from Residential Choice Experiments”. Se- lected paper prepared for presentation at the American Agricultural Economics As- sociation Annual Meeting, Orlando, FL, July 27–29, 2008. Ojea, Elena and Maria L. Loureiro (2011). “Identifying the scope effect on a meta- analysis of biodiversity valuation studies”. Resource and Energy Economics 33 (3), pp. 706–724. Olsen, Søren Bøye et al. (2011). “Tough and Easy Choices: Testing the Inﬂuence of Utility Difference on Stated Certainty-in-Choice in Choice Experiments”. Environmental and Resource Economics 49 (4), pp. 491–510. Ordóñez, Lisa (1998). “The Effect of Correlation between Price and Quality on Consumer Choice.” Organizational behavior and human decision processes 75 (3), pp. 258–273. 180 BIBLIOGRAPHY

Orquin, Jacob L. and Simone Mueller Loose (2013). “Attention and choice: A review on eye movements in decision making”. Acta Psychologica 144 (1), pp. 190–206. Osborne, Jason W. (2010). “Improving your data transformations: Applying the Box-Cox transformation”. Practical Assessment, Research & Evaluation 15 (12), pp. 1–9. Özdemir, Semra et al. (2010). “Who pays Attention in Stated-choice Surveys?” Health economics 19, pp. 111–118. Paish, Oliver (2002). “Small hydro power: Technology and current status”. Re- newable and Sustainable Energy Reviews 6 (6), pp. 537–556. Park, K. M. (2007). “Antecedents of Convergence and Divergence in Strategic Positioning: The Effects of Performance and Aspiration on the Direction of Strategic Change”. Organization Science 18 (3), pp. 386–402. PBL (Netherlands Environmental Assessment Agency) (2016). Trends in global CO2 emissions 2016 report. PBL Netherlands Environmental Assessment Agency. URL: http://edgar.jrc.ec.europa.eu/news_docs/jrc- 2016- trends-in-global-co2-emissions-2016-report-103425.pdf. Peters, Ellen and Paul Slovic (1996). “The Role of Affect and Worldviews as Ori- enting Dispositions in the Perception and Acceptance of Nuclear Power”. Journal of Applied Social Psychology 26 (16), pp. 1427–1453. — (2007). “Affective asynchrony and the measurement of the affective attitude component”. Cognition & Emotion 21 (2), pp. 300–329. Phillips, Peter J. and Gabriela Pohl (2014). “Prospect theory and terrorist choice”. Journal of Applied Economics 17 (1), pp. 139–160. Poe, Gregory L., Kelly L. Girarud, and John B. Loomis (2005). “Computational Methods for Measuring the Difference of Empirical Distributions”. American Journal of Agricultural Economics 87 (2), pp. 353–365. Poe, Gregory L., Eric K. Severance-lossin, and Michael P. Welsh (1994). “Mea- suring the Difference (X - Y) of Simulated Distributions: A Convolutions Ap- proach”. American Journal of Agricultural Economics 76 (4), pp. 904–915. Ponce, Roberto D. et al. (2011). “Estimating the Economic Value of Landscape Losses Due to Flooding by Hydropower Plants in the Chilean Patagonia”. Water Resources Management 25 (10), pp. 2449–2466. Post, Thierry et al. (2008). “Deal or No Deal? Decision Making under Risk in a Large-Stake TV Game Show and Related Experiments”. American Economic Review 98 (1), pp. 38–71. BIBLIOGRAPHY 181

Prognos AG (2012). Die Energieperspektiven für die Schweiz bis 2050. Basel. URL: http://www.bfe.admin.ch/php/modules/publikationen/stream. php?extlang=de&name=de_564869151.pdf. Puckett, Sean M. and David A. Hensher (2008). “The role of attribute processing strategies in estimating the preferences of road freight stakeholders”. Trans- portation Research Part E: Logistics and Transportation Review 44 (3), pp. 379– 395. — (2009). “Revealing the extent of process heterogeneity in choice analysis: An empirical assessment”. Transportation Research Part A: Policy and Practice 43 (2), pp. 117–126. Rabin, Matthew (1998). “Psychology and Economics”. Journal of Economic Litera- ture 36 (1), pp. 11–46. Reisen, Nils, Ulrich Hoffrage, and Fred W Mast (2008). “Identifying decision strategies in a consumer choice situation”. Judgment and Decision Making 3 (8), pp. 641–658. Rigby, Dan, Michael Burton, and Jo Pluske (2016). “Preference Stability and Choice Consistency in Discrete Choice Experiments”. Environmental and Resource Eco- nomics 65 (2), pp. 441–461. Ringquist, Evan J. (2013). Meta-Analysis for Public Management and Policy. San Francisco: Jossey-Bass. Ritenour, Amber E. et al. (2008). “Lightning injury: A review”. Burns 34 (5), pp. 585–594. Robbins, Jesse Lance and Lynne Y. Lewis (2009). “Demolish it and they will come: Estimating the economic impacts of restoring a recreational ﬁshery”. Journal of the American Water Resources Association 44 (6), pp. 1488–1499. Rogers, R. W. (1983). “Cognitive and physiological processes in attitude change: A revised theory of protection motivation”. Social Psychophysiology. Ed. by J. Cacioppo and R.E. Petty. New York: Guilford, pp. 153–176. Rosch, R. (1975). “Cognitive reference points”. Cognitive Psychology 7, pp. 532– 547. Rose, John M. and Michiel C. J. Bliemer (2013). “Sample size requirements for stated choice experiments”. Transportation 40 (5), pp. 1021–1041. Rose, John M. et al. (2008). “Designing efﬁcient stated choice experiments in the presence of reference alternatives”. Transportation Research Part B: Methodolog- ical 42 (4), pp. 395–406. 182 BIBLIOGRAPHY

Rosenberger, Randall S. and Tom D. Stanley (2006). “Measurement, generaliza- tion, and publication: Sources of error in beneﬁt transfers and their management”. Ecological Economics 60 (2), pp. 372–378. Rulleau, Bénédicte and Jeanne Dachary-Bernard (2012). “Preferences, rational choices and economic valuation: Some empirical tests”. Journal of Socio-Economics 41 (2), pp. 198–206. Ryan, Mandy and Verity Watson (2009). “Comparing Welfare Estimates from Payment Card Contingent Valuation and Discrete Choice Experiments”. Health economics 18 (11), pp. 389–401. Samuelson, Paul A. (1937). “A Note on Measurement of Utility”. The Review of Economic Studies 4 (2), pp. 155–161. San Miguel, Fernando, Mandy Ryan, and Mabelle Amaya-Amaya (2005). “’Ir- rational’ stated preferences: A quantitative and qualitative investigation”. Health Economics 14 (3), pp. 307–322. Sandman, Peter M., Neil D. Weinstein, and Paul Miller (1994). “High Risk or Low: How Location on a "Risk Ladder" Affects Perceived Risk”. Risk Analysis 14 (1), pp. 35–45. Saqib, Najam U. and Eugene Y. Chan (2015). “Time pressure reverses risk preferences”. Organizational Behavior and Human Decision Processes 130, pp. 58–68. Sawe, Nik (2017). “Using neuroeconomics to understand environmental valuation”. Ecological Economics 135, pp. 1–9. Sawe, Nik and Brian Knutson (2015). “Neural valuation of environmental resources”. NeuroImage 122, pp. 87–95. Scarpa, R., M. Thiene, and David A. Hensher (2010). “Monitoring Choice Task Attribute Attendance in Nonmarket Valuation of Multiple Park Management Services: Does It Matter?” Land economics 86 (4), p. 817. Scarpa, Riccardo, Danny Campbell, and W. George Hutchinson (2007). “Beneﬁt Estimates for Landscape Improvements : Sequential Bayesian Design and Re- spondents’ Rationality in a Choice Experiment”. Land Economics 83, pp. 617– 634. Scarpa, Riccardo et al. (2009). “Modelling attribute non-attendance in choice experiments for rural landscape valuation”. European Review of Agricultural Eco- nomics 36 (2), pp. 151–174. Scarpa, Riccardo et al. (2013). “Inferred and stated attribute non-attendance in food choice experiments”. American Journal of Agricultural Economics 95 (1), pp. 165–180. BIBLIOGRAPHY 183

Schaafsma, Marije et al. (2014). “Temporal stability of preferences and willingness to pay for natural areas in choice experiments: A test-retest”. Resource and Energy Economics 38, pp. 243–260. Schmidt, Ulrich (2003). “Reference dependence in cumulative prospect theory”. Journal of Mathematical Psychology 47 (2), pp. 122–131. Schulte-Mecklenbeck, Michael, Ryan O. Murphy, and Florian Hutzler (2011). “Flash- light - Recording information acquisition online”. Computers in Human Behav- ior 27 (5), pp. 1771–1782. SED (Swiss Seismological Service) (2016). Basel 1356. URL: http://www.seismo. ethz . ch / de / knowledge / earthquake - country - switzerland / historical-earthquakes/basel-1356/. Seiler, Michael J. and Kimberly F. Luchtenberg (2014). “Do Institutional and Indi- vidual Investors Differ in Their Preferences for Financial Skewness”. Journal of Behavioral Finance 15 (4), pp. 299–311. SFOE (Swiss Federal Office of Energy) (2012). Wasserkraftpotential der Schweiz. Ab- schätzung des Ausbaupotenzials der Wasserkraftnutzung im Rahmen der Energies- trategie 2050. URL: http://www.bfe.admin.ch/themen/00490/00491/ index.html?lang=de&dossier_id=00803. — (2013). Energieperspektiven 2050. Zusammenfassung. URL: http://www.bfe. admin.ch/energiestrategie2050/index.html?lang=de. — (2016). Schweizerische Elektrizitätsstatistik 2015. URL: http : / / www . bfe . admin.ch/php/modules/publikationen/stream.php?extlang= de&name=de_816097016.pdf&endung=SchweizerischeElektrizitatsstatistik2013. — (2017). Elektrizitätserzeugung- und verbrauch 2016. URL: http://www.bfe. admin.ch/energie/00588/00589/00644/index.html?lang=de& msg-id=66433. SFSO (Swiss Federal Statistical Office) (2014). Spezifische Todesursachen. URL: https: //www.bfs.admin.ch/bfs/de/home/statistiken/gesundheit/ gesundheitszustand/sterblichkeit-todesursachen/spezifische. html. — (2016). Statistik der Bevölkerung und der Haushalte (STATPOP). URL: https: //www.bfs.admin.ch/bfs/de/home/statistiken/bevoelkerung/ erhebungen/statpop.html. Shannon, Claude E. (1948). “A mathematical theory of communication”. The Bell System Technical Journal 27, pp. 379–423. Slovic, Paul (1987). “Perception of risk”. Science 236 (4799), pp. 280–285. 184 BIBLIOGRAPHY

Slovic, Paul and Elke U. Weber (2011). “Perception of risk posed by extreme events”. Regulation of Toxic Substances and Hazardous Waste. 2nd Edition. Ed. by John S. Applegate et al. New York: Foundation Press. Slovic, Paul et al. (2007). “The affect heuristic”. European Journal of Operational Research 177 (3), pp. 1333–1352. Smith, V. Kerry and Laura L. Osborne (1996). “Do contingent valuation estimates pass a “scope” test? A meta-analysis”. Journal of Environmental Economics and Management 31 (3), pp. 287–301. Smith, V. Kerry and Subhrendu K. Pattanayak (2002). “Is meta-analysis a Noah’s ark for non-market valuation?” Environmental and Resource Economics 22 (1), pp. 271–296. Soliño, Mario et al. (2012). “Generating electricity with forest biomass: Consis- tency and payment timeframe effects in choice experiments”. Energy Policy 41, pp. 798–806. Sooﬁ, Ehsan S. (1994). “Capturing the Intangible Concept of Information”. Jour- nal of the American Statistical Association 89 (428), pp. 1243–1254. Spinks, Jean and Duncan Mortimer (2016). “Lost in the crowd? Using eye-tracking to investigate the effect of complexity on attribute non-attendance in discrete choice experiments”. BMC Medical Informatics and Decision Making 16:14, pp. 1–13. Stathopoulos, Amanda and Stephane Hess (2012). “Revisiting reference point formation, gains-losses asymmetry and non-linear sensitivities with an em- phasis on attribute speciﬁc treatment”. Transportation Research Part A 46 (10), pp. 1673–1689. Sternberg, R. (2008). “Hydropower: Dimensions of social and environmental co- existence”. Renewable and Sustainable Energy Reviews 12, pp. 1588–1621. — (2010). “Hydropower’s future, the environment, and global electricity systems”. Renewable and Sustainable Energy Reviews 14 (2), pp. 713–723. Stommel, Evelyn (2013). Reference-Dependent Preferences. Wiesbaden: Springer Gabler. Stoutenborough, James W., Shelbi G. Sturgess, and Arnold Vedlitz (2013). “Knowl- edge, risk, and policy support: Public perceptions of nuclear power”. Energy Policy 62, pp. 176–184. Sullivan, Kathryn and Thomas Kida (1995). “The Effect of Multiple Reference Points and Prior Gains and Losses on Managers’ Risky Decision Making”. Organizational Behavior and Human Decision Processes 64 (1), pp. 76–83. Sundqvist, Thomas (2002). “Power Generation Choice in the Presence of Envi- ronmental Externalities”. PhD thesis. Lulea University of Technology, Lulea, BIBLIOGRAPHY 185

Sweden. URL: https://pure.ltu.se/portal/files/153854/LTU- DT-0226-SE.pdf. Sundt, Swantje and Katrin Rehdanz (2015). “Consumer’s willingness to pay for green electricity: A meta-analysis of the literature”. Energy Economics 51, pp. 1– 8. Susaeta, Andres et al. (2011). “Random preferences towards bioenergy environmental externalities: A case study of woody biomass based electricity in the Southern United States”. Energy Economics 33 (6), pp. 1111–1118. Swait, Joffre and Wiktor Adamowicz (2001a). “Choice Environment, Market Com- plexity, and Consumer Behavior: A Theoretical and Empirical Approach for Incorporating Decision Complexity into Models of Consumer Choice”. Orga- nizational Behavior and Human Decision Processes 86 (2), pp. 141–167. — (2001b). “The Inﬂuence of Task Complexity on Consumer Choice : A Latent Class Model of Decision Strategy Switching”. Journal of Consumers Research 28 (1), pp. 135–148. Swait, Joffre and Jordan J. Louviere (1993). “The Role of the Scale Parameter in the Estimation and Use of Multinomial Logit Models”. Journal of Marketing Research 30, pp. 305–314. Tabi, Andrea and Rolf Wüstenhagen (2017). “Keep it Local and Fish-Friendly: Social acceptance of hydropower projects in Switzerland”. Renewable and Sus- tainable Energy Reviews 68, pp. 763–773. Tanaka, Yutaka (2004). “Major Psychological Factors Determining Public Accep- tance of the Siting of Nuclear Facilities”. Journal of Applied Social Psychology 34 (6), pp. 1147–1165. Thurstone, L. L. (1927). “Psychophysical Analysis”. The American Journal of Psy- chology 38 (3), pp. 368–389. Train, Kenneth E. (2009). Discrete Choice Methods with Simulation. New York, NY: Cambridge University Press. Tsuge, Takahiro, Atsuo Kishimoto, and Kenji Takeuchi (2005). “A choice experiment approach to the valuation of mortality”. Journal of Risk and Uncertainty 31 (1), pp. 73–95. Tversky, Amos and Daniel Kahneman (1991). “Loss Aversion in Riskless Choice: A Reference-Dependent Model”. Quarterly Journal of Economics 106 (4), pp. 1039– 1061. — (1992). “Advances in Prospect-Theory - Cumulative Representation of Uncer- tainty”. Journal of Risk and Uncertainty 5 (4), pp. 297–323. 186 BIBLIOGRAPHY

Uggeldahl, Kennet et al. (2016). “Choice certainty in Discrete Choice Experi- ments: Will eye tracking provide useful measures?” Journal of Choice Modelling 20, pp. 35–48. Van Houtven, George, Melonie B. Sullivan, and Chris Dockins (2008). “Cancer premiums and latency effects: A risk tradeoff approach for valuing reductions in fatal cancer risks”. Journal of Risk and Uncertainty 36 (2), pp. 179–199. Van Loo, Ellen J. et al. (2014). “Visual Attribute Non-Attendance in a Food Choice Experiment: Results From an Eye-tracking Study”. Agricultural & Applied Eco- nomics Association’s 2014 AAEA Annual Meeting, Minneapolis, MN, July 27-29, 2014. Viscusi, W. Kip (1998). Rational Risk Policy. Oxford: Oxford University Press. Viscusi, W. Kip and Joel Huber (2012). “Reference-dependent valuations of risk: Why willingness-to-accept exceeds willingness-to-pay”. Journal of Risk and Uncertainty 44 (1), pp. 19–44. Viscusi, W. Kip, Joel Huber, and Jason Bell (2014). “Assessing whether there is a Cancer Premium for the Value of a statistical Life”. Health economics 23, pp. 384–396. Visschers, Vivianne H. M., Carmen Keller, and Michael Siegrist (2011). “Climate change beneﬁts and energy supply beneﬁts as determinants of acceptance of nuclear power stations: Investigating an explanatory model”. Energy Policy 39 (6), pp. 3621–3629. Waechter, Signe, Bernadette Sütterlin, and Michael Siegrist (2015). “Desired and Undesired Effects of Energy Labels—An Eye-Tracking Study”. Plos One 10 (7), e0134132. Wakker, Peter P. (2010). Prospect Theory For Risk and Ambiguity. Cambridge UK: Cambridge University Press. Wang, Hua (1997). “Treatment of "don’t know" responses in contingent valuation surveys: a random valuation model”. Journal of Environmental Economics and Management 32 (2), pp. 219–232. Warren, Charles R. et al. (2005). “’Green On Green’: Public Perceptions of Wind Power in Scotland and Ireland”. Journal of Environmental Planning and Man- agement 48 (6), pp. 853–875. Weber, Elke U. and Richard A. Milliman (1997). “Perceived risk attitudes: relating risk perception to risky choice”. Management Science 43 (2), pp. 123–144. Weber, Martin and Colin F. Camerer (1998). “The disposition effect in securities trading: an experimental analysis”. Journal of Economic Behavior & Organiza- tion 33 (2), pp. 167–184. BIBLIOGRAPHY 187

Weisser, Daniel (2007). “A guide to life-cycle greenhouse gas (GHG) emissions from electric supply technologies”. Energy 32 (9), pp. 1543–1559. Weller, Priska et al. (2014). “Stated and inferred attribute non-attendance in a design of designs approach”. Journal of Choice Modelling 11 (1), pp. 43–56. Whitﬁeld, Stephen C. et al. (2009). “The future of nuclear power: Value orienta- tions and risk perception”. Risk Analysis 29 (3), pp. 425–437. Willemsen, Martijn C. and Eric J. Johnson (2008). MouselabWEB Documentation. URL: http://www.mouselabweb.org/download.php. — (2010). MouselabWEB. URL: http://www.mouselabweb.org/. Winer, R. S. (1986). “A reference price model of brand choice for frequently pur- chased products”. Journal of Consumer Research 13, pp. 250–256. Witte, Kim (1992). “Putting the Fear back into Fear Appeals: The Extended Par- allel Process Model”. Communication Monographs 59, pp. 329–349. Wolsink, Maarten (2000). “Wind power and the NIMBY-myth: institutional ca- pacity and the limited signiﬁcance of public support”. Renewable Energy 21 (1), pp. 49–64. Wüstenhagen, Rolf, Maarten Wolsink, and Mary Jean Bürer (2007). “Social acceptance of renewable energy innovation: An introduction to the concept”. Energy Policy 35 (5), pp. 2683–2691.