Estimating certainty, risk and time effects on the expected utility for energy contracts

(1)

Estimating certainty, risk

and time effects on the

expected utility for energy

contracts

Using conjoint analysis with stated probabilities

Emma Turkenburg

(2)

1

Title page

Estimating certainty, risk and time effects on the expected utility for energy contracts Using conjoint analysis with stated probabilities

Master thesis Emma Turkenburg Moesstraat 65 9741AA Groningen +31648631848 e.s.turkenburg@gmail.com s1995502 Faculty of Economics and Business Rijksuniversiteit Groningen 1st supervisor: Keyvan Dehmamy 2nd supervisor: Felix Eggers 20-6-2016

(3)

2

Preface

I would like to express my gratitude to Professor Dehmamy, my research supervisor, for his patient guidance and useful critiques of this research work. I would like to offer my special thanks to all my friends and family that took the time to fill out my questionnaire, making it possible to write this thesis. Also I would like to thank my fellow thesis students for the moral support that was needed to keep working. Finally, I would like to thank my family and friends for their support and their

unwavering believe in my abilities.

Management Summary

(4)

3

1. Introduction

1.1 Theoretical problem

There has already been a long tradition of research into the effects of time, uncertainty and risk on the utility of consumers. The measurement of preference parameters under conditions of time, uncertainty and risk and their underlying psychological mechanisms have become increasingly important (Ida and Goto, 2009). Methods have already been developed to simultaneously measure time and risk preferences. By using discrete choice experiments Rachlin, Rainer and Cross (1991), Keren and Roelofsma (1995), Anderhub, Güth, Gneezy and Sonsino (2001), Yi, de la Piedad and Bickel (2006) and Ida and Goto (2009) have measured time and risk preferences.

The problem with all these experiments is that the underlying psychological mechanisms of time and risk are often linked to a feeling of uncertainty that the consumer feels when making the choices presented in these experiments. However, not all of this uncertainty is due to the person's inability to exactly state what his or her time and risk preferences are. Part of this uncertainty is due to the fact that in all these experiments the choice scenarios are incomplete. Respondents are given only a subset of the information they would have in real choice settings. By doing so stated choices might differ from actual choices, making these stated choices nothing more than point predictions of actual choices (Manski, 1999).

It should be an aim of future research to take away as much of this uncertainty as possible. One way of assessing part of this uncertainty is by giving respondents the opportunity to state their uncertainty about their behaviour within incomplete choice scenarios. This can be done by using conjoint analysis with elicited choice probabilities. Eliciting choice probabilities has also some drawbacks. Respondents tend to round of the probabilities that they state, also asking respondents to state probabilities instead of their most preferred choice increases the cognitive burden that is placed on these respondents (Manski ,2004; Manski and Molinari, 2010; Blass, Lach, and Manski 2010). The relevancy of these drawbacks will be addressed in this thesis by comparing the method of eliciting probabilities to the method of stated choices.

The method of eliciting probabilities was first introduced by Manski (1999). Furthermore, a first successful empirical implementation was performed by Blass et al. (2010). This paper will empirically evaluate this new method by directly comparing the results of stated choices and elicited probabilities on energy contracts.

(7)

6 presented, followed by a discussion on the social and theoretical relevance of posing and answering these questions. The rest of the structure of this paper will be discussed at the end of this introductory chapter.

1.2 Background energy contracts

In this thesis conjoint analysis based on elicited probabilities will be used to estimate risk, time and uncertainty effects on preferences for consumer energy contracts. The Dutch energy market has gone through severe changes in the past decade. These changes might have had an effect on the consumers’ view on energy contracts and the preferences that they have among the several options that are available in the market. Especially, since the choices among contracts have become increasingly complex since the liberalization of the energy market.

A little over a decade ago, in 2004, the energy market in the Netherlands was liberalized. This liberalization entails that, where before the energy supplier and ownership of the energy delivery network was combined into one company, now they have become separate companies. This separation has made it possible for consumers to freely choose their energy supplier, while before this depended on where a person lived. Now, a decade later some major changes have taken place. Since 2004 consumers have been able to switch suppliers, also they have gone from six possible suppliers to over thirty suppliers (energieleveranciers.nl, accessed: march, 2016).

In a comparable study by Blass et al. (2010) on Israeli households’ energy supply, they used the reliability of the energy supply to estimate time and risk preferences. The Dutch energy market however is characterized by a very high level of reliability, in the past 14 years since 2014, the average amount of electricity availability has been 99,995% (ECN, energie-nederland and netbeheer nederland, 2014). Therefore, unlike the study by Blass et al. (2010) the reliability of the energy supply is not used to estimate the effects of risk, uncertainty and time. Rather risk effects reside in choices that can be made between variable or fixed payment rates. Uncertainty is assessed based on the respondents’ confidence about their own knowledge. Furthermore, time effects are assessed by looking at choices between contract duration.

(8)

7 energy prices might drop causing a consumer to lose money when choosing a fixed tariff. However, energy prices are hard to predict since there are many factors that influence the price (vereniging eigen huis, accessed 2016). By using conjoint analysis with elicited probabilities it will become more clear how customers deal with risk, time and uncertainty effects when estimates about total prices and usage are taken away.

Adding to the confusion for customers are the choices for contract duration. Choices for contract duration are somewhat counterintuitive. There is a gap between what is best for customers and what is best for the firms. Customer loyalty has many marketing advantages for firms such as reduced marketing costs, more new customers, greater trade leverage, favourable word of mouth and greater resistance among loyal customers to competitive strategies (Dick and Basu, 1994; Chaudhuri and Holbrook, 2001). Customers are also aware that firms want them to be loyal, therefore it would be logical to assume that it is beneficial to stay for longer periods of time with the same supplier. But as competition has become fierce, firms battle for new customers. Many of them offer welcoming presents, often in the form of cash payments on the bill of the first year of energy supply. As these cash payments are a one-time occurrence per contract, it becomes significantly cheaper for customers to switch suppliers every year (energieprijzenvergelijken.com; accessed 2016). Even though this is the case, ever since switching has become possible only 44% of the Dutch consumers has ever changed energy suppliers and in 2013 only 12% of the consumers switched. Consumers who have never changed energy contracts pay on average €350, -- more per year than consumers who change contracts every year (ECN, energie-nederland, netbeheer nederland, 2014). Using conjoint analysis with elicited choice probabilities this thesis will estimate the time preference of consumers regarding their energy contract duration.

All in all, it becomes more and more difficult for consumers to have a clear and certain order of preferences. Also, when consumers become uncertain about their preferences, it becomes more difficult for suppliers to predict their buying behaviour and meet the needs of these customers. This thesis will attempt to give insight into preferences of consumers for energy contracts, especially focussing on how time, risk and uncertainty which are inherent to these choices influence the consumers’ preferences.

1.3 Problem statement

(9)

8 uncertainty effects on expected utilities. This will be done by applying both methods (stated choices and elicited choices) to conjoint surveys on energy contracts. The energy market has been changing heavily, making it difficult for customers to be certain about their preferences, in turn making it difficult for the suppliers to cater to the needs of the customers. The uncertainty that customers experience when having to make choices about attributes such as contract duration and variable vs. fixed tariffs could have an effect on the expected utilities of these customers. This thesis will give more insight on how uncertain customers form expectations about their utilities when being faced with choices that vary in time and risk. From a marketing perspective this might be interesting for practitioners, such as marketing managers, involved in offering customers’ products or services that vary in time and risk such as energy contracts.

1.4 Research questions

In this thesis the following research questions will be answered:

● Is there empirical proof to show that conjoint analysis with elicited probabilities is more statistically relevant than conjoint analysis with stated choices to estimate time and risk and uncertainty effects for expected utilities?

● Do consumers’ risk preferences and uncertainty about choices for energy contracts influence expected utilities for certain types of energy contracts?

● How do consumers’ risk preferences and uncertainty about choices for energy contracts influence the expected utilities for certain types of energy contracts?

1.5 Theoretical and social relevance

(10)

9 customers. Secondly, the choices that have to be made between fixed and variable tariffs and contract duration are inherently linked to risk and time effects. This thesis will attempt to clarify which constructs in energy contracts create uncertainty, resulting in guidelines for both suppliers and customers on how to reduce uncertainty about the choices to be made, making customers more confident about the right choice of contract and giving suppliers a clearer image on the preferences of their customers. This is important as the most important reason for opening up the energy market has been to reduce energy prices (Kirschen, 2003). However, many people refrain from switching energy suppliers, resulting in them paying on average €350, -- more than when liberalization was introduced. Part of this has to do with uncertainty about the choices to be made. In a study under 494 Dutch consumers by the Dutch authority on consumers and markets, it became clear that only a quarter of the respondents felt like the energy suppliers are focussed on what the consumers’ wants and needs are. Even more interesting was that consumers who have not switched energy suppliers since the liberalization of the market are the most mistrustful (ACM, 2015). This problem should be addressed, especially for new players in the energy market this would be an interesting problem to solve. Solving this problem starts by getting more insight into how consumers form expected utilities for energy contracts.

1.6 Structure of thesis

In the upcoming chapter the theoretical framework will be presented, including theory about the difference between eliciting probabilities and stating choices in conjoint analysis, and psychological mechanisms that determine how consumers will form their preferences. With these theories hypotheses are formed which will lead to the formation of a conceptual model. Following this chapter, a description will be given of the methodologies that will be used to collect and analyse the data. Consequently, the results and their validity and reliability, which will be based on these analyses will be presented. After which, a chapter will be included in which the reasons and underlying theories for both expected and unexpected results will be discussed. This thesis will end by giving some general recommendations for theory, practice and future research.

2. Theoretical framework

(11)

10

2.1 Stated choices vs. Elicited probabilities

This subchapter discusses the differences between conjoint analysis with stated choices and conjoint analysis with elicited probabilities. The main difference between these two methods is that with stated choices it is assumed that a respondent in a choice survey is aware of both the utility that he gives to a certain option and the error term. In practice, this is almost never the case because it has been the norm to pose only subsets of information (Blass et al., 2010). As a result of receiving these subsets of information, respondents possibly decide differently in surveys than in reality (Manski, 1999).

In a conjoint analysis model with stated probabilities a respondent 𝑖 is asked to make a choice in several choice scenarios, which each have a number of hypothetical alternatives 𝑗. In most conjoint analyses with stated choice models the utility of alternative 𝑗 will have a random coefficients form of (1) 𝑈𝑖𝑗= 𝑥𝑖𝑗𝛽𝑖+𝜀𝑖𝑗

With stated-choice analysis it is assumed that the respondent 𝑖 who responds to the choice scenario is aware of both the values of 𝑥𝑖and 𝜀𝑖. Through knowing these values the respondent 𝑖 is able to determine with certainty what the utility-maximizing choice is. In reality the researcher only shows the respondent 𝑖 a subset of the characteristics (𝑣𝑖 = (𝑣𝑖𝑗, 𝑗 = 1, … , 𝐽)). These characteristics determine 𝑥𝑖 but not 𝜀𝑖. Therefore it is not very probable that the respondent knows 𝜀𝑖 based on the scenario that was presented to him (Blass et al., 2010).

Conjoint analysis with elicited probabilities will solve many of these problems. As mentioned before one of the problems with conjoint analysis comes from not giving respondents the full set of characteristics that a product would have in reality. By not doing this, the respondent is not able to know the value 𝜀𝑖. Conjoint analysis with elicited probabilities addresses this problem by enabling respondents to express uncertainty about 𝜀𝑖. In this case the respondent can treat 𝜀𝑖 as a vector of utility components whose value might not be known through the characteristics shown in the survey but which would be known when the same choice would have to be made in reality. Formally this means that the respondent 𝑖 would form a subjective distribution for 𝜀𝑖,from this the respondent would derive the subjective probability that he would choose either product in reality and report these to the researcher (Blass et al., 2010).

(12)

11 based on the characteristics 𝑣𝑖 as presented in the survey, the subjective choice probability for alternative 𝑗 will be:

(2) 𝑞𝑖𝑗= 𝑄𝑖[𝑥𝑖𝑗𝛽𝑖+ 𝜀𝑖𝑗> 𝑥𝑖𝑘𝛽𝑖+ 𝜀𝑖𝑘, all 𝑘 ≠ 𝑗]

On the right side of this equation (2) the subjective random utility interpretation of elicited choice probabilities is given. The distribution as formed by the respondent 𝑄𝑖 shows the resolvable uncertainty of the respondent. Resolvable uncertainty stands for the uncertainty about utility that a respondent places on the characteristics that are not included in the conjoint analysis, but that would be known in reality. A respondent can also encounter situations where he faces only unresolvable uncertainty. In these situations, the general economic assumption is that the respondent will base his choice on the maximization of his subjective expected utility. Formally this means that the respondent would state a subjective probability of one for the alternative that would maximize his expected utility (Blass et al., 2010).

Herein lies also the difference between the methods with stated choices and elicited probabilities. With stated choice models there is no distinction between resolvable or unresolvable uncertainty. With these models the reasoning is that a respondent forms his subjective choice probability and simply responds by stating the alternative with the highest probability. The problem is that while the respondent means that the probability of choosing that product in reality is the highest, the researcher analyses the data as if the respondent would state that his utility for that product is the highest (Juster, 1966; Manski, 1990). More formally stated:

The respondent means to state that 𝑞𝑖𝑗 > 𝑞𝑖𝑘 for all 𝑘 ≠ 𝑗 The researcher translates this as 𝑈𝑖𝑗 > 𝑈𝑖𝑘 for all 𝑘 ≠ 𝑗

However, this assumption only holds when the respondent would only experience unresolvable uncertainty (Blass et al., 2010). Sadly, this is generally not the case since most surveys are based on incomplete choice scenarios.

Also the computational issues of having to find the structural parameters of utility by using backward recursion to estimate dynamic choice models are solved by using elicited probabilities (Hotz and Miller, 1993). By using elicited probabilities, the utilities can be expressed in terms of utility payoffs, which are the choice probabilities. For dynamic models this means that it is possible to estimate structural parameters that also hold for future periods (Hotz and Miller, 1993).

(13)

12 Dominitz, Manski, and Heinz (2003). By using the example of social security benefits their research shows that by eliciting probabilities about the respondents’ expectations about their eligibility for social security benefits they were able to show how the amount of certainty that these respondents expressed was spread across the population.

It can be concluded that by eliciting probabilities instead of stating choices much more relevant information about how respondents form expectations about their utility for a certain product is gathered. By using this method, a respondent reporting non-extreme probabilities can express his belief that in reality he would possess more choice-relevant information beyond the characteristics provided in the survey. By reporting extreme probabilities of 1 and 0 the respondent can express that he does not expect to learn anything extra about the product in reality (Blass et al., 2010). As this method gives more information per choice set the amount of information that can be gathered per respondent greatly increases. Therefore, it should be possible to reach higher statistical significance with this method compared to the conjoint analysis method of stated choice. This leads to the first hypothesis:

H1: The conjoint analysis with elicited probabilities has greater statistical significance than the conjoint analysis with stated choices.

2.2 Uncertainty and the development of energy prices

As mentioned in the introduction the two conjoint analysis methods will be compared by applying them to a conjoint study on energy contracts. More information about how these energy contracts are build up and what psychological mechanisms underlie the formation of utilities for these contracts are presented in the following subchapter.

2.2.1 The development of energy prices.

Each energy contract consists of several components. It is common for energy contracts to include both electricity per kilowatt hour and gas per square meter (energieprijzenvergelijken.nl, accessed; 2016). Both the prices for energy and gas consist of the same components. Every component has an effect on the total price that a consumer pays for its electricity bill. The components are;

· Either fixed or variable pay per use tariffs · Contract duration

(14)

13 Energy contracts have stochastic elements incorporated in them. This makes it difficult for consumers to determine the true value of what they are buying, which creates uncertainty. The influence of this uncertainty in combination with how risk averse the consumer is can have differing effects on the expected utility. According to Isik (2006) uncertainty, risk preferences and irreversibility all have an effect on the willingness to pay of a consumer. In the upcoming section the effects of risk preferences, irreversibility through time and uncertainty about the development of energy prices on the utility for each of the components of energy contracts will be discussed.

2.2.2 Contract duration

Choices that involve decisions about time are influenced by certainty effects. The certainty effect refers to the observation that people seem to prefer outcomes that are considered certain relative to outcomes which are merely probable (Kahneman and Tversky, 1979). Keren and Roelofsma (1995) conducted an experiment where the respondents got choices between monetary rewards that varied in time and certainty. It showed that people have a preference for the option that gives immediate utility because options further away in time are always less certain. However, when the utility payoffs of both options are further away in the future the probability of choosing the earlier option reduced by 45%, reversing the order of preferences. Meaning that for options that are both far away in the future the utility for price differences is discounted less heavily. This effect is known as hyperbolic discounting. Hyperbolic discounting is the best documented discounted utility anomaly. In short it entails that people prefer smaller sooner rewards to a larger-later reward. It also shows that the discount rate over longer time horizons is lower than the implicit discount rate over shorter time horizons (Frederick and Loewenstein, 2002). When fitting a mathematical function to these types of choice data a hyperbolic functional form fits the data better than the exponential function, which would imply constant discount rates. It should be noted that all these theories are based on rewards, e.g. the receiving of monetary rewards by waiting for a certain time. In the case of choosing an energy contract the choice is not about the receiving of a monetary reward but about the possible opportunity cost of rising energy prices. As this opportunity cost can still be considered a reward, the expectation is that the effect will be the same. Therefore:

(15)

14

2.2.3 Tariff type

How people form their preferences for the type and amount of tariff will also be influenced by uncertainty effects. Uncertainty in this case relates to how confident people feel about the knowledge they possess on the subject at hand. Metcalfe (1986) and Schacter (1983) distinguish between different types of knowledge. The first type of knowledge is subjective knowledge, which is a combination of knowledge and self-confidence, also referred to as self-perceived knowledge (Park and Leisig, 1981). The other type of knowledge is the level of actual knowledge (Metcalfe, 1986; Schacter, 1983). These different types of knowledge have different effects on decision outcomes such as perceived task complexity and perceived confusion during task performance (Brucks, 1985; Park and Lessig, 1981).

When consumers do not know whether energy prices will fall or rise, both their self-perceived and actual knowledge regarding their choice for energy contracts will be low. This will induce feelings of uncertainty. When a consumer is uncertain about the development of energy prices, variable tariffs might be perceived as a risk. Therefore, it is to be expected that consumers who are risk averse and who are uncertain about the development of energy prices will avoid contracts with variable tariffs. Another effect of this uncertainty is that the consumers who experience this might have a decreased expected utility for longer contracts as this increases the fear of irreversibility (Kahneman and Tversky, 1979). Furthermore, positive time preferences, which is defined as having a preference for things that are of shorter duration or closer in time, are mainly a result of risk aversion, an earlier end date decreases the consumers’ liability which as a result decreases the corresponding uncertainty (Olson & Bailey, 1981). Also Becker et al. (1964) found that there is a statistically significant correlation between an individual’s risk aversion and this person’s discount rate. Consumers with a high degree of risk aversion discount the future more heavily. This leads to the conclusion that individuals who are characterized by a relatively high level of risk aversion will prefer contracts of a shorter duration.

(16)

15 Hypothesis three and four will combine the theories about how utilities are formed for tariff types with theories on how utilities are formed for welcoming presents. Therefore, the hypotheses belonging to this section will be presented at the end of the following subchapter.

2.2.3 Welcoming presents

Uncertainty about the future can take different forms. Even when the future outcome is said to be certain, the individual can perceive uncertainty. He can perceive uncertainty about the outcome, meaning there is no way to assess whether you will receive what was promised. Also, he can perceive uncertainty about himself, for example by having doubt about being in a situation where you are able to collect what was promised. Because of this a future outcome’s utility is discounted. Furthermore, research has shown that when an actual risk of not receiving what was promised is added, the discounting is increased (Benzion, Rapoport and Yagil, 1989). Laibson (1997) showed that this issue can only be resolved through immediate consumption. Experimental evidence showed that there is an immediacy effect in choice behaviour. Respondents tend to choose earlier smaller rewards over later larger rewards when this earlier reward offers immediate utility. As people who are uncertain about the development of energy prices and risk averse are expected to be more focussed on immediate consumption, they might be more focussed on welcoming presents.

In the paper by Isik (2006) this idea is confirmed, investigating willingness to pay (WTP) for goods that give uncertain pay-offs in terms of utility, he finds that WTP decreases as uncertainty increases. However, this WTP rises if risk averse people are compensated for the risk that they take, as part of the irreversibility of the risk that consumers take will be compensated. As WTP is directly related to the expected utility that consumers give to a product it might be probable that respondents experiencing internal uncertainty and who do not want to take risks will prefer more immediate pay-offs such as larger welcoming presents.

For people who are uncertain about the development of prices, and who are risk averse there are opposing forces within energy contracts. On the one hand, the shortest contract is the riskiest, while a long contract leads to a greater uncertainty due to irreversibility. Furthermore, uncertain consumers may want to feel like they are compensated for the risk that they are taking, therefore they will want a welcoming present.

This leads to the following hypotheses:

(17)

16 H4: Consumers who are uncertain about the development of energy prices will prefer fixed tariff contracts that are (A) relatively short (1 or 2 years) and (B) larger welcoming presents as compensation for the risk that they perceive tot take.

2.2.4 Fixed supply costs

The uncertainty that comes from a lack of knowledge might also relate to the amount of energy that an individual uses. This type of uncertainty could influence the preference for the fixed supply costs of an energy contract. In studies on telephone contracts and health club contracts (Kling and Van der Ploeg, 1990; Kridel, Lehman and Weisman, 1993; Train McFadden and Ben-Akiva, 1987; Della Vigna and Malmendier, 2006; Nunes, 2000) it came to light that there are consumers who can save money with a pay-per-use tariff who will still choose a flat rate. This phenomenon is known as the ‘flat-rate-bias’ (Train, 1991). There are several reasons for having such a bias that relate to the choice of energy contracts. First, Risk averse consumers may prefer larger fixed costs to avoid unexpected variation in the yearly bill. In a sense they are insuring themselves against unexpected costs. When customers are risk averse they are expected to want this insurance because they cannot predict if there will be periods where they will use more than average (Miravete, 2002; Train, 1991; Winer, 2005). With energy contracts this can occur due to extra cold or longer than expected winter periods.

Choosing higher fixed costs might also be a result of mere convenience (Winer, 2005). When consumers feel overloaded by information they seem to have a preference for flat rates. There is a bias towards flat rates when consumers are unknowledgeable about the cost differences between pay per use or flat rates. This is also due to consumers not wanting or knowing how to calculate which option is more expensive. Therefore, although a flat-rate might not be cheaper, it is more clear to the consumer what he needs to pay.

With energy contracts it is not possible to choose a completely flat rate. There is always a combination of a base rate and pay-per use tariff. No literature is available on how utilities are formed when flat rates and pay per use rates are combined into one choice. However, as long as a higher base rate leads to a clear decrease in the pay per use tariff, the result will still be that more variation due to pay per use tariffs will be taken away. This is what decreases the uncertainty as described in the above section. Therefore, the arguments as presented here are expected to hold for base rates in energy contracts as well. Consequently:

(18)

17 H6: Consumers who are uncertain about their usage of energy will prefer higher base rates under the condition that this leads to a decline in the pay per use tariff.

2.2.5 Expecting energy prices to drop

For consumers who are indifferent about taking risks or consumers who feel certain about energy prices dropping different mechanisms are important. In essence uncertainty about a products true value negatively influences the willingness to pay for that product. This effect can be mediated by the amount of risk preference a person has. Meaning that if the person is indifferent about the risk that he is taking, the negative effect on utility might be less (Isik, 2006).

It might also happen that consumers feel certain that energy prices are going to drop. In that case the utility of the energy contract will become larger over time. In most experiments about discounting there are two choices. e.g. a smaller amount of money now, or a larger amount of money later. In these experiments positive discounting is the norm (Frederick and Loewenstein, 2002). However, when experiments are performed where respondents have to choose between sequences of pay-offs, they seem to prefer improving sequences of pay-offs over declining sequences of pay-offs (Ariely and Carmon, 2002; Frederick and Loewenstein, 2002; Loewenstein and Prelec, 1993). This concept also works the other way around. There have also been experiments that showed people prefer declining discomfort over sequences where the discomfort stays equal (Varey and Kahneman, 1992).

With energy contracts, there is the possibility to choose between variable and fixed tariffs. Variable tariffs can be beneficial to the customer when energy prices drop over time. Contracts with variable prices are always freely terminable. It is quite difficult to predict how energy prices will develop as there are many factors influencing these prices (vereniging eigen huis; accessed 2016). This leads to the conclusion that consumers who expect energy prices to drop need to also be somewhat indifferent about risks in order to prefer variable tariffs. As people have a preference for improving sequences (Frederick and Loewenstein, 2002) they will be less focussed on immediate rewards such as welcoming presents. Therefore:

H7: If a risk indifferent consumer expects energy prices to drop they will have a preference for (a) contracts with variable tariffs without contract duration and (b) low welcoming presents.

(19)

18

2.2.6 Churning

As mentioned before the consumers are expected to have a positive (hyperbolic) time preference. This means that it is to be expected that most people will want shorter contracts rather than longer contracts. This preference for shorter energy contracts has already been confirmed in a study by Goett (1998). However, the figures about the current energy market show a different picture. 49% of the consumers has never switched energy suppliers, in 2013 only 12% percent churned, in 2015 this amount increased to 14% (ACM, 2015). This is the case even though consumers who have never switched pay on average €350, -- more than consumers who switch every year. There are two possible explanations for this discrepancy. First of all, buying a new product or service is inherently uncertain, whereas the current product or service can feel comfortable as the consumer is already familiar with its features. This effect can be so strong that consumers are willing to trade in some of their price sensitivity for the convenience of staying with the same familiar product or service (Aydin, Özer and Arasil, 2005). Related to this is the existence of switching costs. Which entails that it takes effort to switch from one supplier to the other. Theses switching costs prevents some customers from churning. This allows firms to charge higher prices, reduce product or service quality or obtain abnormal returns (Klemperer, 1987; Lieberman and Montgomery, 1988; Farrell and Klemperer, 2007). Due to these mechanisms:

H9: There will be a discrepancy between the indicated positive time preference and the amount of churn among respondents. (people will state that they want to switch every year, even though they don’t actually do this)

2.3 Conceptual model

Hypotheses were formed based on the theories that have been presented. These theories and the hypotheses lead to the formation of a general conceptual model, which is presented in figure 1. In the conceptual model the components of which an energy contract consists are shown. By default, it is expected that people will have a negative utility for the base rate and for the amount of pay per use tariff, since these increase the overall price of the energy contract as they become larger. The utility for the welcoming present will be positive as this lowers the price of the energy contract. The choice for either a fixed or variable contract will completely depend on the risk averseness and expectations of the consumer. Only consumers who are indifferent about risks or consumers who expect energy prices to drop will have a preference of variable tariffs. As uncertainty or risk averseness increases people are expected to opt for short fixed tariff contracts.

(20)

19 welcoming presents. Being uncertain about the amount of energy usage and risk aversion will increase the utility for fixed supply costs.

Conceptual Model

(21)

20

3. Research design

In order to measure the hypotheses as proposed in the theoretical framework a survey will be used. This survey will ask respondents to indicate their preferences regarding different types of energy contracts. The contract will have several attributes such as contract duration, cash welcoming presents, either a fixed or variable tariff to pay for use and fixed supply costs. Contract duration and fixed vs. variable tariffs will be combined into one attribute as they are confounded with each other in reality as well. Each choice set will contain four possible contracts. The survey will be designed in a specific way to accommodate the analyses of all hypotheses. Furthermore, other subjective data will be collected to enrich the data (Manski, 2004). In the upcoming section the survey design will discussed in more detail per hypothesis.

3.1 Research methods

For the first hypothesis, conjoint analysis with stated choices will be compared to conjoint analysis with elicited probabilities. This will be achieved by having the respondents first answer seven questions in which they will have to indicate their most preferred choice. In the next seven questions the respondents will be asked to state how probable it would be that they would buy each option. For hypothesis two, the utilities for contract duration need to be measured in such a way that the difference between closer time horizons and time horizons which are further away in the future can be measured. This means that in the seven questions with elicited probabilities, there will be choice sets containing short time horizons and longer time horizons. In order to be able to do this at least five levels need to be included. The freely terminable contract is linked to variable tariffs, which means that this could skew the results. Because of this only the utilities across the fixed tariff contracts may be compared to each other. Therefore, there is a need for at least one option which is variable in price and freely terminable, and to compare across time horizons there is a need for at least four levels with fixed tariffs. This need determines that there will be five levels for the attributes. To prevent a number of levels effect, the other attributes will also have five levels each (Eggers, 2015).

3.1.1 Collection of additional information

(22)

21 knowledgeable they feel about energy contracts and their risk averseness. It is possible to measure risks averseness using self-reported information on risky behaviour (Dohmen, Falk, Huffman, Sunde, Schupp and Wagner, 2011). The study by Dohmen et al. (2011) investigates different measures of risk behaviours. They have found that all the self-reported information that was given by the respondents was related to several behaviours, which was proof of their behavioural validity. Overall the best explanatory variable was the general risk question, which had predictive validity for all behaviours. However, for each specific context a question relating to that context is also a good predictor of that specific behaviour. In the case of this thesis, the risky behaviour is applied in a financial context. Therefore, two statements will be posed to determine the respondent's degree of risk averseness. One general statement;

1. I general, I am willing to take risks And one specific financial statement;

2. I would invest my money in risky financial products

In order to measure how certain people feel about their knowledge on the development of energy prices and energy usage the following statements will be included in the survey;

3. I am able to make informed decisions about energy contracts

4. I have enough knowledge about energy prices to know how they are going to develop 5. I am aware of the amount of energy that I use

6. I have enough knowledge about my use of energy to make informed decisions about energy contracts

Hypothesis seven and eight concern people their expectations regarding the development of energy prices. The questions in the survey will concern only gas prices, an explanation for this will follow in the next subchapter. Also important to note is that the sliders that will be used will range from completely disagree to completely agree, the middle value will indicate that the respondent is neutral. The following statements will be included to control for the difference between respondents filling out neutral for this statement because they do not know what is going to happen and filling out neutral because they think energy prices will stay the same;

7. Gas will become more expensive in the near future 8. Gas prices will stay the same in the near future

(23)

22 9. I should switch energy suppliers every year

10. I switch energy suppliers every year

The first statement should clarify whether people are aware that switching is profitable for them. The second statement should clarify whether people actually do switch every year. These statements will be compared to each other and to the utility that people indicate for contract duration.

The respondents will be asked to reply to these statements in the form of a slider, indicating to what degree they agree or disagree with the statements. By doing so, the answers to the statements will be recorded as continuous data, which makes it possible to use them in a continuous mixture model, and which decreases the number of covariates that have to be estimated.

3.1.2 Attributes and levels

In the upcoming section the attributes and levels that will be included in the conjoint analysis survey will be discussed in more detail. Also some other issues concerning the design of the conjoint study will be addressed.

(24)

23 Distribution of Energy Contract Attributes

Minimum 25th percentile 50th percentile 75th percentile Maximum Variable Tariffs Gas €0.57 €0.57 €0.58 €0.63 €0.65 Fixed Tariffs Gas €0.50 €0.54 €0.57 €0.58 €0.63 Fixed Supply Costs €0. -- €29. -- €42. -- €73. -- €102. -- Welcoming Present €0. -- €0. -- €14.50 €50. -- €260. -- Contract Duration

0 year 1 year 1 year 3 years 5 years

Total Price €961.92 €1064.69 €1116.85 €1151.92 €1459.24

Table 1

As the number of attributes should be manageable, most experiments include less than seven attributes. Using more attributes can result in choices that are too complex for the respondents to answer (Green, Srinivasan, 1978; 1990). Most energy contracts consist of both an offering for gas and electricity. The way these two commodities are offered in terms of their prices is similar, they both include a pay per use tariff and the other attributes such as contract duration, the welcoming present and the fixed supply costs are either split across the two commodities or equal amounts for both commodities. Because of the great similarities between the way the two commodities are offered, including both the gas and the electricity offerings with all their attributes would overcomplicate the survey. Also adding these extra attributes will not provide any extra explanatory power, since the psychological mechanisms discussed in the literature review influence these attributes in the same way. Therefore, this survey will only include attributes concerning gas prices. Gas has been chosen as the commodity to include in the survey because gas is still the biggest cost of energy in the average household. In 2013, 38% of the average household’s energy usage was through gas, opposed to 28% through electricity (energietrends, 2014).

(25)

24 terminable contracts. To prevent for example a two-year variable tariff contract from being included in the survey, the type of tariff (variable vs. fixed) and contract duration will be combined into one attribute. By doing so, no attributes will be interrelated with each other.

Taking these issues into account, it can be concluded that there should be four attributes included in the survey. These attributes and their corresponding levels are presented in table two.

Attributes and Levels Levels/ Attributes Welcoming present Fixed Supply Costs

Contract duration & tariff type

tariff price

1 €0.-- €0.-- freely terminable and

variable tariffs

€0.45

2 €15.-- €30.-- 1 year and fixed tariffs €0.50

3 €50.-- €60.-- 2 years and fixed tariffs €0.55

4 €150.-- €90.-- 4 years and fixed tariffs €0.60

5 €250.-- €120.-- 6 years and fixed tariffs €0.65 Table 2

There are also some issues to be considered regarding the number and specification of the levels. First of all, the levels should have a greater range than what they are in reality, while still being acceptable. By doing so you are able to cover possible future developments (Green and Srinivasan, 1978; 1990; Teichert, 2001). The ranges of the levels are all based on the figures as presented in table one. They are all somewhat wider than they are in reality, except for the welcoming present. The welcoming present is an exception, as the maximum value that was found was an extreme outlier compared to most contracts, which becomes clear by comparing the 75th percentile (€50. --) to the maximum value (€260. --) in table one.

Another important issue to consider is that the number of levels should be kept to a minimum, since the complexity of the experimental design increases exponentially as the number of levels is increased. As mentioned before minimum number of levels that is required for this study is five. Five levels are needed because both the variable tariff, and four different years of fixed contracts are needed to be able to test all the proposed hypotheses.

(26)

25 balanced across all attributes. Otherwise, results can become skewed because the attributes with more levels will artificially become more relevant to the consumers. Therefore, each attribute has five levels, as can be seen in table two.

3.1.3 Factorial design

In essence, full factorials are always preferred, since they are always balanced and orthogonal, meaning that the attribute levels each occur an equal number of times and that each pair of attribute levels is balanced (Eggers, 2015). However, in this thesis testing every potential combination is not possible as that would mean that 5^4 = 625 possible combinations, would have to be tested. However, a full factorial is only required if all main effects and all potential interaction effects have to be estimated. Since we are not interested in all interaction effects for this thesis, a fractional factorial design will be used. This fractional design will be created by using Sawtooth software.

The efficiency of the fractional design can be tested easily by checking the output of the test design in Sawtooth. If there are no or only minor correlations between the attributes then the design is orthogonal and the parameters can be identified without bias (Eggers, 2015). The results of these tests will be presented in as a separate paragraph at the end of this sub chapter.

3.1.4 Choice design

(27)

26 In this thesis there will be four stimuli per choice set, this number creates a sound balance between gathering as much information as possible per choice set, while keeping the amount of information presented to the respondent manageable (Batsell and Louviere, 1991). Balanced overlap can simply be achieved through Sawtooth functionality. Furthermore, no dominated choice sets will be shown by indicating for each attribute level the natural order. The only exception will be contract duration and tariff type, because the natural order of preference is not clear for this attribute since the value of each level depends on price developments in the market. For example, if prices rise over four years, then having a four year fixed contract would be most beneficial, if on average prices decline over four years, a freely terminable contract with variable tariffs would be most beneficial.

3.1.5 Survey design

In the following section the design of the survey will be presented, showing what the questions will look like, and what has been done to make the survey as clear and easy to use as possible for the respondents. Pictures of the survey can be found in appendix A.

The first page of the survey will give a short introduction, welcoming the respondent, and giving a short explanation of what is expected of the respondent. Also some demographic information such as age and gender will be asked of the respondent before continuing with the rest of the survey. On the next page the statements with sliders will be shown. The respondents will be asked to rate the statements, ranging from strongly disagree to strongly agree, the middle being neutral.

After having responded to the statements, a more detailed introduction will be given to the first seven choice sets. In this introduction the respondents will be told that they have to indicate which would be their most preferred option if these were their only options of energy contracts. Furthermore, some basic information about the characteristics of energy contracts are given to prevent confusion.

The first seven choice sets will be presented, which will be choice sets asking the respondents to state their choices.

Next an introduction will be given for the third part of the survey in which respondents will be asked to state how probable it would be that they would choose each option in reality. A picture of these instructions and an example of a probability choice task are also included in appendix A.

3.1.6 Testing the design

(28)

27 tasks, which each include four concepts and four attributes. In table three the a priori estimates of the standard deviations for the attribute levels are shown. From this table it becomes clear that the design is minimally different from optimal as only the last two levels of the attribute welcoming present are not shown an equal number of times. It is also important to interpret the actual and ideal standard deviations. These measures are approximations of the relative standard deviations of each main effect under aggregate analysis, assuming that each version of the questionnaires is unique across all observations. The test design uses ordinary least squares instead of multinomial logit to do this. The focus of this test is on the relative magnitudes between the standard deviations. The actual standard deviation gives the standard deviations based on the data file, the ideal standard deviations on the other hand give the standard deviations for when the design would be precisely orthogonal with the same number of observations (Sawtooth manual, Accessed: 23-05-2016). The biggest difference recorded here is 0.0002, this value is low enough to be ignored. The efficiency measure in the last column of table three is a squared ratio of the design in terms of estimating each parameter relative to the hypothetical orthogonal design). Overall these values are all very close to one, which is the desired value. Therefore, it can be concluded that the design is orthogonal.

In order to test how many respondents are needed more advanced tests are needed. These advanced tests simulate random respondent answers to the questionnaire. By doing so it is assumed that respondents are heterogeneous with unknown preferences. Using this simulated data set the test performs an aggregate logit which estimates the effects that have been selected, in the case of this questionnaire only main effects are considered. The design was tested several times in order to find the appropriate amount of respondents that have to be gathered in order to achieve significant results. In general, the standard errors should not exceed 0.05 (Sawtooth manual, Accessed 23-05-2016). As only the standard error for the attribute level of €90. -- welcoming present is higher than 0.05 (0.06) the design is acceptable when 230 respondents are collected. This is based on a none option percentage of one percent. The reason for choosing one percent is that 98% of the Dutch households are connected to gas supply (ACM, 2015), as the sample consists of mainly highly educated people and students who live in the city, the two percent of the population that does not want to be connected to gas will probably not be captured in the sample of this study.

3.2 Data collection

(29)

28 needed to answer hypotheses three until nine) are included at the end of the questionnaire also a shorter version excluding the choice exercise was created.

Respondents have been targeted through social media such as Facebook and LinkedIn also personal emails have been send. Furthermore, the survey has been promoted using the survey website A Priori Estimates of Attributes and Levels

Att/Lev Levels Freq. Actual SD Ideal SD Difference SD Effic.

1/1 0 year 1120 Deleted 1/2 1 year 1120 .0453 .0452 .0001 .9973 1/3 2 years 1120 .0452 .0452 0 1.0006 1/4 4 years 1120 .0453 .0452 .0001 .9984 1/5 6 years 1120 .0452 .0452 0 1.001 2/1 €0.45 1120 Deleted 2/2 €0.50 1120 .0455 .0453 .0002 .9902 2/3 €0.55 1120 .0454 .0453 .0001 .9967 2/4 €0.60 1120 .0452 .0453 -.0001 1.0026 2/5 €0.65 1120 .0451 .0453 -.0002 1.0095 3/1 €0.-- 1120 Deleted 3/2 €30.-- 1119 .0452 .0453 -.0001 1.0027 3/3 €60.-- 1120 .0452 .0453 -.0001 1.0039 3/4 €90.-- 1121 .0454 .0453 .0001 .9947 3/5 €120.-- 1120 .0454 .0453 .0001 .9954 4/1 €0.-- 1120 Deleted 4/2 €15.-- 1120 .0453 .0452 .0001 .9965 4/3 €50.-- 1120 .0451 .0452 -1E-04 1.005 4/4 €150.-- 1121 .0451 .0452 -1E-04 1.0053 4/5 €250.-- 1119 .0454 .0452 .0002 .9945 Table 3

(30)

29 will increase the amount of respondents. Their might however be a lack of motivation to give relevant answers as the aim of the respondent might be to get to the end of the survey as quickly as possible, this will be further discussed in the results chapter.

3.3 Plan of analysis

In the upcoming subchapter the plan of analysis will be discussed. The steps that have been taken to reach conclusions about the hypotheses will be discussed in chronological order.

The data that was collected through the survey has been analysed using statistical software from Sawtooth. Additional analyses have been performed in SPSS.

As the testing of the first hypothesis will be conducted using less respondents than the other hypotheses the description of the dataset as used for the first hypothesis will be discussed separately. After discussing the demographics and characteristics of this dataset the extra respondents which will be used to assess the other hypotheses will be added and discussed.

The data has been cleaned by removing the respondents that gave to little information and by assessing the motivation of the respondents. Respondents that show a lack of motivation could interfere with the validity of the results therefore they should be deleted from the dataset (Wise and Kong, 2005). The motivation of the respondent is assessed by looking at the response time and at respondents who answer all questions in the same way. Respondents who are not motivated will show rapid guessing behaviour, which means that they will answer at a rate that makes it impossible to even consider the entire question (Wise and Kong, 2005). The results of assessing these issues for both datasets can be found in appendix B.

From the 107 respondents that have finished the choice exercise there are twelve people who have not finished the probability exercise. In order to compare the two exercises twelve respondents from the short survey that will be used to test the other hypotheses have been added to the dataset. In order to get an initial feeling for the data, some of the statements that will be used as covariates will be investigated.

(31)

30 In order to control for the effect that different questionnaires might have on the results, a covariate will be included in the models that will indicate which version of the questionnaire the respondent filled out, other control variables that are available for estimation are age and gender. For each covariate that relates to the testing of the hypotheses two to four items were included in the questionnaire. Therefore, the covariates will first be combined using factor analysis to prevent multicollinearity among the covariates. The principal component method was used with Varimax rotation.

The models have been created in an evolutionary way. First estimating a basic model including only the attributes as part-worth parameters, after which the models will be made more complex by adding covariates, linear parameters or interaction effects. For hypothesis one, multiple models for both the choice and the probability exercises have been created in order to compare. For both exercises four models have been created. Respectively each exercise has a model without any covariates or linear parameters, a model including all covariates without any linear parameters, a model in which the parameters that could be linear are estimated as linear, and an optimal model based on the previous model estimations.

For both the choice models and the probability models the utilities have been plotted to see whether the parameters should be estimated as linear.

For both exercises there are control covariates and explanatory covariates available. The control variables were added to both exercises to increase reliability.

The performance of the models will be assessed using the information criteria AIC and BIC. Furthermore, the model fit criteria that were produced by Sawtooth will also be assessed, which are the percentage certainty, the root likelihood, the parameter RMS and the average variance. Bayesian factors are also important model fit criteria to consider (Kass and Raftery, 1995). The Bayesian factor takes into account the prior probabilities. Bayesian factors are closely related to the BIC or Schwarz criterion. The BIC gives a rough approximation of the logarithm of the Bayes factor. Because the Bayesian factor considers the prior probabilities it provides information on the models uncertainty. In simple terms the Bayesian factor can be calculated as follows;

𝐵𝑎𝑦𝑒𝑠𝑖𝑎𝑛 𝐹𝑎𝑐𝑡𝑜𝑟 = 𝑃𝑜𝑠𝑡𝑒𝑟𝑖𝑜𝑟 𝑜𝑑𝑑𝑠 𝑃𝑟𝑖𝑜𝑟 𝑜𝑑𝑑𝑠

(32)

31 The first model fit criterion is the percent certainty by Hauser (1978). It indicates how much better the model is than a chance model compared to the perfect solution. The percent certainty can be calculated in the following way:

𝑃𝑒𝑟𝑐𝑒𝑛𝑡 𝐶𝑒𝑟𝑡𝑎𝑖𝑛𝑡𝑦 =𝐿𝐿(1) − 𝐿𝐿(0) −𝐿𝐿(0)

The outcome varies between zero and one, where one means a perfect fit, while zero means the model only fits at chance level.

The next model fit criterion is the Root Likelihood (RLH), it is quite similar to the percentage certainty measure. The RLH is calculated by taking the 𝑛th root of the likelihood, where n is the total number of choices made by all respondents in all tasks. In this case 𝑛 is 107 ∗ 7 = 749. In other words, the RLH is the geometric mean of the predicted probabilities. If there would be no information about the part-worths, each alternative would be chosen with a probability of 1/𝑘, where k is the number of alternatives per choice set, which is 0.2. If the RLH is one the fit is perfect (Sawtooth manual, accessed 24-05-2016).

The third model fit is the average variance. This is a less direct model fit measure. The scaling of the part-worths depends on the goodness of fit, meaning that the better the fit, the larger the estimated parameters. The average variance is the average of the current estimate of the variance of part-worths across respondents. In other words, if the range of the attribute utilities becomes larger so does the model fit (Sawtooth manual, accessed 24-05-2016).

The final model fit criteria that is provided in the output of Sawtooth is the parameter root mean square (RMS). It is related to the average variance. It gives the root mean square of all part worth estimates, across all part-worths and over all respondents (Sawtooth manual, accessed 24-05-2016). In order to compare the models, the AIC and BIC need be computed. They have been derived in the following manner; the log likelihood of the null model was calculated:

𝐿𝐿(0) = 𝑁 ∗ 𝑃 ∗ ln (1 𝑐)⁄

Where 𝑁 is the number of respondents, 𝑃 is the number of tasks in the questionnaire and 𝑐 is the number of concepts per task. The likelihood of the null model is -1205.47. Using the percentage certainty as mentioned before, the log likelihoods of the models have also been calculated.

The AIC is calculated according to the following function: 𝐴𝐼𝐶 = 2𝑘 − 2ln (𝐿)

(33)

32 The BIC is calculated according to the following function:

𝐵𝐼𝐶 = −2 ln(𝐿) + 𝑘 ∗ ln (𝑛) Where 𝑛 is the number of observations.

The significance of the utilities has been assessed using the Wald test. The Wald-statistic is used because it performs the best for maximum likelihood estimations with small sample sizes (Leeflang et al., 2015).

The Wald statistic has been calculated to test whether the ranges of the individual utilities are significantly different from zero. Normally the Wald test to compare two values requires the covariance between these two values. Using this covariance decreases the joint standard error as without this covariance it is assumed that these standard errors can move in random directions compared to each other. Therefore, not having this covariance makes this test more conservative. For the linear attributes the statistic was calculated to test whether the individual utility slopes are significantly different from zero.

The linear parameters have been calculated using the following function: Wald statistic_ia

=

𝑈𝑖𝑎−𝑈0

𝑆𝐸(𝑈𝑖𝑎)

Where U0 has been set to zero as the statistic is used to test whether the slope is significantly different from zero.

The part-worth estimates have been calculated using the following function: Wald statistic_ia

=

max(𝑈𝑖𝑎)−min (𝑈𝑖𝑎)

√𝑆𝐸_{max(𝑈𝑖𝑎)}2 +𝑆𝐸_{min (𝑈𝑖𝑎)}2

Where Uia is the preference of respondent (i), for attribute (a) expressed as a utility (U).

(34)

33 The covariates are calculated by taking the averages over the final 10,000 iterations, the significance is determined by whether the estimates have a consistent sign, 90% or 95% of the time (Orme and Howell, 2009).

4. Results

4.1 Hypothesis one

4.1.1 Demographics

For testing the first hypothesis 153 respondents were recorded. These respondents were quite equally distributed based on gender, as 51.6% was male. The distribution of age is given in figure 2 Looking at the distribution it becomes clear that most of the respondents are concentrated around 23 years old. This is to be expected as most respondents were gathered from direct social networks such as Facebook. Inspecting the graph further it becomes clear that there is another peak in age between fifty and sixty years old, this is also to be expected as a large number of respondents were collected through secondary social networks such as family. Furthermore, it becomes clear that the youngest person to fill out the questionnaire was 17, while the oldest respondent was 71. The 12 respondents that were added from the questionnaire including only probability do not significantly change any of the characteristics.

Distribution of Age

Figure 2

4.1.3 Characteristics

In table four the descriptives are presented of the items that will be included in the model as covariates. Looking at the first statements that will be used as covariates it becomes clear that overall, people tend to be quite moderate when it comes to risk aversion (57.56), but when it comes

(35)

34 to financial products, the respondents are a lot more risk averse (29.07). Close inspection of the statements about the ability to make decisions (58.81), the knowledge about price developments (32.50) and the knowledge (47.27) and awareness on energy usage (50.01) all seem to be quite evenly spread across the sample, the boxplots that show this can be found in appendix C.

Descriptives of Covariates

Minimum Maximum Mean

General risk 0 100 57.56

Risky products 0 99 29.07

Ability to make decisions 0 100 58.81

Knowledge about prices 0 100 32.50

Aware of energy usage 0 100 50.01

Knowledge about usage 0 100 47.27

Energy prices rising 0 100 55.79

Energy prices stay the same 0 88 39.32

Switch every year 0 100 24.18

Should switch every year 0 100 47.82

Table 4

A cross examination of age and these statements can be found in table five, it shows that overall people tend to get more confident about their knowledge as they grow older, therefore it might be wise to include an interaction effect between age and these variables.

Correlations between Age and Certainty Items Bivariate Correlations Ability to make decisions Knowledge about prices Aware of Energy Usage Knowlege about Energy Usage Age .324 .425 .445 .437 Table 5

4.1.5 Factor analysis

(36)

35 that it is useful and possible to do factor analysis. The solution provides four factors, there appear to be no cross loadings. The results can be found in table six. The first factor includes the ability to make decisions, the knowledge about price developments, the awareness of energy usage and the

knowledge about energy usage, this factor will be named ‘Certainty’. The second factor includes general risk taking and investing in financially risky products, this factor will be named ‘Risk Taking’. The third factor includes the statements about energy prices rising or energy prices staying equal, these statements will be called ‘Price Rises’. The last factor includes the statements about if the respondent has switched or thinks he/she should switch energy suppliers; this factor will be named ‘Churning’.

Factor Analysis

Items/Factors Certainty Risk Taking Price Rises Churning

General Risk .005 .865 -.075 .059

Risky products .136 .812 .118 .057

Ability to make decisions .808 -.074 -.082 .113

Knowledge about price development .762 .292 .117 .088

Aware of energy usage .858 .054 .005 .048

Knowledge about energy usage .914 .005 .002 .010

Energy prices will rise -.124 .196 -.818 -.028

Energy prices will stay the same -.104 .284 .730 .024

Switch every year .290 .144 -.247 .754

Should switch every year -.048 .007 .255 .858

KMO .690

Bartlett’s Test 354.606**

Table 6

** Significant at 99% confidence level

Estimating certainty, risk and time effects on the expected utility for energy contracts