Colour Priming and
Risk Behaviour
Experimental approach by the use of BRET
Loenen, W.H.T. (Wout), s4355075 15-8-2018
Supervisor: dr. Jana Vyrastekova
Master: Economics, Behaviour, and Policy Abstract
This study is one of the first that investigates the relationship between colour and risk behavior. Behavioural economists has shown in the past that behavior can be
changed by all sort of small interventions. The one still missing and heavily understudied is the influence of colour on behaviour. Three colour are used; grey, red, and green. The literature on red is substansive in contrast to the literature on green and grey is the control colour. However, risk behavior is understudied. In this
study the experimental approach is utilized by the use of BRET (Crosetto & Filippin, 2014). This provides the opportunity to isolate the colour effect. The result show a
slight effect of colour priming on risk behavior between red and green but not between the control task and the treatment tasks. This provides a stepping stone
for further research to see if this effect is persistent. However, the control questions show that financial risk behavior is not in line with actual risk behavior and the gap between the revealed and stated preferences still has to be resolved.
2
Table of Content
1. Introduction. ... 3
2. Theoretical Framework ... 5
2.1 Financial decision making ... 5
2.1.1. Uncertainty and Risk ... 6
2.2 Priming ... 8
2.3 Colour Research... 10
2.2 Risk Behaviour Assessment. ... 13
2.2.1 Risk Elicitation Tasks ... 13
2.2.2 Comparison ... 15 2.4 Control questions ... 18 2.5 Hypotheses. ... 19 3. Experimental design ... 20 3.1 Colour use. ... 21 3.2 Experiment pay-off. ... 21 3.3 Empirical Method. ... 22 3.3.1 Methodology. ... 24 3.3.1.1Dependent Variables. ... 24 3.3.1.2 Independent variables. ... 24 3.4 Method. ... 25 4. Results ... 26 5. Discussion ... 32 6. Conclusion ... 36 7. Literature ... 37 8. Appendix ... 42 Appendix 1 ... 43 Appendix 2 ... ... 46 Appendix 3 ... 47 Appendix 4 ... 48 Appendix 5 ... 49 Appendix 6 ... 50 Appendix 7 ... 51
3
1. Introduction.
Without gaining much attention in the literature, colours are widely present in the financial decision making process. Examples of such are: advertisements online; television reports; stock market reflections and at firm’s websites. The colours most prominently employed are red and green. However, the literature mostly neglected the associations made by seeing colour and how this affects the financial decision making process. Economists are mostly interested in other aspects that can influence the decision making process such as risk and uncertainty. Recently, behavioural economics has gained ground and provided insight in influencing the decision making by the use of nudges. This is contradicting the traditionalist economic view of the rational agent that posses the skill of unlimited information access and an infinite cognitive capacity to make the decision what would result in the highest obtained utility. The traditional economic view implicitly assumes that the agent would not be affected by any subtle cue or intervention as this rational agents could cognitive comprehend the goal of the seller to influence its decision. So, any intervention in this sense would be rendered useless in the traditional view.
However, behavioural economics has provided evidence that behaviour can be changed in various ways and is affected by little changes made in the decision making process. The presence of colour in the decision making process can be placed under ‘priming’. Priming can be defined as an unconscious remembering process when certain stimuli occur, the brain processes this as additional information to make the decision, which is an important aspect of social thought (Kliger & Gilad, 2012a). Another definition of priming is that people’s actions be altered by exposure to certain sights, words, or sensations (Dolan et al., 2012). To give an insight on how much visual appearance matters; from 11 million bits of information processed in our brain per second, 10 million of these bits are going through our visual system (Wilson, 2002). It would be ignorant to leave colour out the equation when analyzing financial behaviour. Consumer mostly have a big spectrum of choices where the relative visual appeal often determine the consumer preferences and decisions (Shi, Baba, & Rao, 2014). A colour can evoke different associations depending on the setting it appears in. Some field studies found that women whom wear red dresses or lipstick were more often approached by males at the bar, red cars made driver behave more aggressive, and consumers eat less food from a red plate(Guéguen, 2012; Bruno, Martani, Corsini, & Oleari, 2013; Guéguen, Jacob, Lourel, & Pascual, 2012). These studies point out that subtle colour cues can affect behaviour in the real world. Colours can potentially alter decisions as they sights can be depicted in colour. To dive deeper in the
4 behaviour, the main goal of this study is to contribute to this literature of colour priming and
financial risk behaviour. In the continuation of this study, the implications of colour priming and financial risk behaviour are discussed to answer the following research question:
Is there a difference in risk behaviour due to the presence of colour priming?
In the literature with the aim on colour priming and risk behaviour the emphasis was mainly on the colour red and how this affected the risk behaviour. Although, risk behaviour is not explicitly researched in context to colour priming, achievement and avoidance behaviour are the most
commonly aspects. This research adds to literature that a direct link between colour priming and risk behaviour is pursued. As financial risky decisions can nowadays be taken at home by the use of internet, colour priming is more prevalent than in the situations as the office of, for example, a bank employee discussing a mortgage.
The method mostly used in these studies is a survey with queues in red or depicting the stock market changes in red and subsequently see if the investment behaviour changes to a more risk averse manner(Kliger & Gilad, 2012a). In this study a simple experiment is conducted to isolate the colour priming and see if this has an effect on risk behaviour. This gives the best insight if colour priming has indeed an effect due to its nature of revealed preferences, the actions subjects show when an alteration is made to an task (P. Crosetto & Filippin, 2016). In the chapters following, the theory about financial decision making, priming, colour and risk measurement are discussed. Following the theoretical chapter, the experimental set-up is set out, the results are discussed and subsequently a conclusion and discussion are there to close the study.
5
2. Theoretical Framework
To formulate hypotheses regarding the research question put forward in section 1, the theoretical and empirical literature is discussed in this section. The section is structured as follows. Firstly, the financial decision making process is set out and linked with decision making under the conditions of uncertainty and risk. Secondly, the aspect of priming in decision making processes is portrayed. Thirdly, the associations and subtle cues people perceive when a colour is added into the decision making process. Lastly, the choice of which task is used to assess the research questions, by comparing the four mostly used risk elicitation task in the literature.
2.1 Financial decision making
The concepts, theories and models of traditional finance and economy assume rationality of persons and efficiency of markets when investigate certain issues. However, recent research in behavioural economics and finance provided contradictory evidence against the assumptions of the rational models(Nigam, Srivastava, & Kumar Banwet, 2018). The underlying assumptions of the rational agents are that they have infinite cognitive capacity and access to real-time information so they can always choose the optimal action. Behavioural economics challenges the fully rational decision maker on the premise that a decision maker’s rationality is bounded (Simon, 1955).
Following the thought of behavioural economists perspective, many non-financial traits can influence the financial decision making process which were previously ignored by the traditional economists. Nigam et al. (2018) provide a set of reasons why non-rationality of agents is real. An example of such is January effect; the average returns in the month January are higher than both the preceding and succeeding moths. This can be attributed to the behavioural false hope syndrome that the new year will result in more prosperity (Anderson, Gerlach, & DiTraglia, 2007; Ciccone, 2011). No economic or financial consideration is taken into the equation or can rationally explain such phenomena.
Moreover, increasing agreement within the behavioural sciences says our behaviour and decision making are strongly influenced by contextual and situational factors we encounter (Dolan et al., 2012). From the traditional economic point of view, this should never occur as the agents can process any information and make the most optimal decision each time. These findings of the behavioural sciences provide more evidence that this is not the case.
At first glance, it seems that decision making should not be the most difficult thing to
understand. However, the behavioural sciences provided substantial evidence that our decisions can be altered by interventions which we consciously notice and still our choice is altered. Furthermore, unconscious interventions which we do not even notice can still affects the decision and actions we make. An example of such unconscious intervention is; participants were asked to make a sentence out scrambled words such as fit, lean, active and athletic and subsequently were more likely to use
6 the stairs instead of the elevator (Wryobeck & Chen, 2003). From the traditional standpoint, no one should take the stairs as the input requires more effort that take the elevator. So with less input the same results is achieved. However, this simple experiment of using words in a sentence can shape the behaviour of the participants. Almost all economic and financial decisions are characterized by a degree of uncertainty and risk (Dohmen et al., 2011). To continue on the questions raised in the introduction, we narrow down the decision making process to a subpart of the spectrum, namely risk behaviour. Almost every financial decision we make is characterized by a certain degree of risk and we investigate which subtle cues can alter the financial decision made by participants.
2.1.1. Uncertainty and Risk
Risk and uncertainty play a role in almost every economic and financial decision. The difference between risk and uncertainty lies within the information the agent has. Uncertainty is when the probabilities of the outcome are not known and thus ambiguous. Risk is when the probabilities are known so the agents can better way the alternatives. Understanding individual attitudes towards risky behaviour is closely linked to understanding and predicting economic and financial
behaviour(Dohmen et al., 2011). Risk can be defined as Jessor (1998) stated: ) risk taking is the engagement in a behaviour that could result in a potential gain although it is balanced by the
potential negative consequences from that same behaviour. The negative consequences here are not related to losses otherwise we are dealing with loss aversion in addition which is beyond the scope of the research objective. In this study, the ‘negative’ consequence is not earning money with the task. A more accurate definition for the field of economics is provided by Harrison & Rutström(2008): risk preference more often refers to the tendency to engage in behaviours or activities that involve higher variance in returns, regardless of whether these represent gains or losses, and is often studied in the context of monetary payoffs involving lotteries. An advantage of the definition of Harrison & Rutström (2008) is that risk behaviour can be tested with merely positive returns and thus provides an solution to isolate risk behaviour from loss aversion.
There have been multiple studies that showed empirical evidence that risk attitudes have a large impact on a scale of important outcomes and behaviour such as educational attainment, home ownership, patterns of occupational choice, wealth, and investment in stocks (Barsky, Juster, Kimball, & Shapiro, 1997; Bonin, Dohmen, Falk, Huffman, & Sunde, 2007; Dohmen et al., 2011). Therefore, it is important to understand if risk behaviour can be altered or not. The illustration of elevator showed that simple behaviour can be altered by subtle cues like words, but does this also apply to risk behaviour? Traditional economics would not agree, as agents are immediately aware of the risk and uncertainty present when acquiring an asset or service and can subsequently precisely assess if it is worthwhile to purchase of the asset or service. Behavioural economics would agree that behaviour
7 can be altered by subtle cues. However, most of the behavioural research regarding risk behaviour is in the field of personal traits agents have. Such personal traits can be certain biases a person posses such as self-serving bias, overconfidence and deposition bias or can reflect gender or ethnic
differences.1 Despite these findings, the use of words and images to change risk behaviour is heavily understudied.
The next step is to see how risk can be measured and if different kinds of risk measurement result in the same overall score or that risk differs individually between different domains such as financial, health or recreational activities? The two most prominently used measurement of risk are stated and revealed preferences. Stated preferences are self-report measures seek to elicit
preferences in response to hypothetical or real-world behaviours, or both. Examples of such are that respondents are asked to rate themselves on scale with opposing poles being “not willing at all to take risks” and “very willing to take risks” or let the subject express their likelihood of engaging in risky behaviour such as: “How likely would you go sailing when it recommend not to due to heavy weather conditions?”(Mata, Frey, Richter, Schupp, & Hertwig, 2018). Revealed preferences are mostly utilized by well-controlled experiments to isolate a selected situational variable has an effect on behaviour using objective measures as outcomes of interest. An example of such is a dictator game where the dictator sees if the receiver has the same gender or not, and see if this deviates from the results where nothing is known about the receiver(Mata et al., 2018).
Both method have its advantages and disadvantages, stated preferences seem to be stable indicators of risk preferences but it is harder isolate a single effect or determinant. Revealed
preferences has the advantage that a single determinant can be isolated to see to what extent it will result in changed behaviour. Both kinds of preferences have been subject to criticism. The lack of generalizability across revealed preferences elicitation methods is reason for concern(Friedman, Isaac, James, & Sunder, 2014). Furthermore, stated preferences has received a substantial amount of scepticism, in both economics (Beshears, Choi, Laibson, & Madrian, 2008) and psychology (Haeffel & Howard, 2010), as this can represent not much more than ‘cheap talk’ (Mata et al., 2018). In this study, a single alteration is made so a revealed preference approach fits the needs the best to answer the research question. However, the stated preferences are not ignored as stated preferences are asked to see if the results found in revealed preferences part have external
validation. Furthermore, Mata et al. (2018) state that scholars of each method almost use exclusively one of the two measurements of preferences. The stated preferences are not the interest of the
1 See, for example (Barsky et al., 1997), (Guiso & Paiella, 2004, 2008), (Donkers, Melenberg, & Van Soest, 2001)
8 study but merely function to see if the results found any external validation. Furthermore, it colour priming could be used in a stated preference approach but it would be less prominent present thus making it harder to answer the research question.
For the purpose of the study, to see if colour priming has any effect on risk behaviour, revealed preferences suits the research question best. As we can isolate a certain specification of the research and colour can be better implemented in the design without the subjects knowing what the research is about, leading to a lower chance of a ‘demand effect’ (Zizzo, 2010). The experimenter demand effect refers to changes in behaviour by subjects due to cues about what constitutes the behaviour expected or wished by the experimenter (Zizzo, 2010). It can take two forms according to Zizzo (2010); cognitive or social. Cognitive demand effect is that a subject indentifies the task presented and behaves accordingly, by employing cues about what behaviour is appropriate for the task. Social demand effect can take two kinds of relationships, namely the experiment-subject and the subject-subject relationship. The first relationship is that the experimenter, explicitly or implicitly, extort social pressure on the subject through instructions and cues. The other relationship is the subject-subject relationship. This occurs when subjects interact with other subjects during the experiment, resulting in conformation to the actions other make during the course of experiment (Zizzo, 2010). This is the so-called social peer pressure. Within this research, the demand effect can only be cognitive by the subjects itself and/or the experimenter-subject social pressure if the experimenters instructions or cues pushes the subject towards a certain direction.
2.2 Priming
It is established that certain changes in the decision making process can affect the decision. There are some different interventions that can act in the conscious realm of the mind but also outside of the conscious realm. Outside of the consciousness means that the interventions are not intentionally observed although the brain processes these subtle cues. In order to answer the research question of this study, we need an intervention that acts outside of the conscious realm. Dolan et al. (2012) provide different methods to influence behaviour unconsciously and the most suitable is priming as this primarily acts outside of the conscious realm. However, affect can also be of effect. As our emotional associations can powerfully shape our actions(Dolan et al., 2012). This can have an influence when people see certain colour or colour expressions but this is discussed in the
subsequent section. For the purpose of the research, priming is used as the behavioural influencing tool. Priming is defined as: an unconscious remembering process when certain stimuli occur, the brain processes this as additional information to make the decision, which is an important aspect of social thought (Kliger & Gilad, 2012). Cramer (1968) defined priming as changes in preliminary
9 definition of Cramer (1968) fits into the research as the preliminary conditions of the task are
changed to obtain a certain response of the stimuli. Different stimuli can be used to alter behaviour. The following definition divides stimuli in three distinct forms; priming is that people’s actions be altered by exposure to certain sights, words, or sensations (Dolan et al., 2012). So there are three different kinds of priming we can use to influence the risk behaviour; words, sensations, and sights. Each of the three is briefly discussed and one is argued to be the best suited of the aim of the research.
First, the use of words as a priming tool is discussed. Words that can evoke certain associations people made with the context these words are most commonly used or can evoke associations that people assign to these words. Example of such is when people are exposed to words relation to elderly people, such as wrinkles, resulted in that walked more slowly out of the room than they walked in and had a poorer memory of the room itself (Dijksterhuis & Bargh, 2001). The example mentioned in the previous section regarding the use of stairs instead of the elevator is another example of how words with specific associations can shape behaviour. The same applies when words such as trust, share, collaborate and teamwork, are primed before a public goods game and significantly increases the contributions to the public good (Drouvelis, Metcalfe, & Powdthavee, 2010). This is not only the case by words that are just written down, it can also occur by just asking people about what they intend to do as such questions that can recall old behaviours or mentally represent new behaviours more easily. An simple intuitive example of such was demonstrated by Levav & Fitzsimons (2006). They asked the participant to indicate the likelihood of flossing their teeth in the coming week and this resulted in a significant increase of the flossing frequency of the
participants.
Second, sensations as a tool is discussed and it is often defined as smell. Some smells can be appropriated for simple things like leaving buildings more clean. Holland, Hendriks, & Aarts (2005) found evidence for that just the scent of all-cleaner led significantly more people leaving their tables clean after eating. Even smell can be in the unconscious realm of the mind, as (Li, Moallem, Paller, & Gottfried, 2007) showed. As they exposed the subjects to a smell (either pleasant, neutral or
unpleasant) below the threshold of detection and the subjects ratings of the likeability of faces they subsequently saw changed according to the smell they were exposed to.
Third, sights as a tool is discussed. Sights can be everything from a smiley to pictures of running shoes and even just normal objects like container structures. Wansink & Kim (2005) primed their subjects with a bigger container for their popcorn at the cinema and found evidence that when primed with a bigger container, subjects significantly ate more. A depiction of eye brows above an
10 honesty box where people can get a drink makes them pay three times more than in the situation where no eye brows are depicted (Bateson, Nettle, & Roberts, 2006). Even money can be used to prime subjects. (Vohs, Mead, & Goode, 2006) used Monopoly money or screensavers showing money to prime their subjects. The subjects were less willing to help another voluntarily, preferred to work alone, donated less to charity and selected leisure experiences that were more individually focused. Their findings were related to relational economic exchange and self-sufficiency.
It is less well understood of all the primes we encounter every day, which have a significant effect and which do not (Dolan et al., 2012). Therefore, it is important to first isolate the single primes that may have an effect on behaviour. Consequently, when having an effect on behaviour in an isolated setting, one can blend in more primes to investigate which primes have dominance above others. So, first we isolate one prime to see if this has any effect on risk behaviour. Moreover, the best fitting prime for implementing colour and answering our research questions is sights. With sights we can prime the object itself or surroundings in the colour of choice, what gives us a set of options. Smells are the least applicable as colour cannot be smelled. Words was the other option. In this case the words can be depicted in the colour of choice. However, this can cause some problems as this may already induce the ‘demand’ effect as subjects can detect this diversion from the normal black font.
2.3 Colour Research
Research in the field of colour influences on behaviour is limited to a few empirical and theoretical studies (Elliot & Maier, 2007). This is peculiar as colour are present in every day interactions regardless of the nature of the interaction or the decision that has to made. In the field of physics the research regarding colour is a subpart of the light and wavelengths, which is beyond the scope of this research. Although some general insights from this field can be of use as why certain colours are not associated with the same objects. For example, objects of colour are never described as reddish-green/greenish-red or bluish-yellow/yellowish-blue. However, other
combinations with these four colours are readily perceived, like green and blue, green and yellow, red and blue, and red and yellow (Fairchild, 2005). Red and green (or yellow and blue) cannot exist together as these hue perceptions are encoded in a bipolar fashion by our visual system(Fairchild, 2005). A hue is; ‘attribute of a visual sensation according to which an area appears to be similar to one of the perceived colours: red, yellow, green, and blue, or to a combination of two of
them.’(Fairchild, 2005). Colours can be divided in two different sets, achromatic and chromatic colours, namely grey, black, and white and blue, yellow, red, and green. As Fairchild (2015) puts it: an achromatic colour is ‘perceived colour devoid of hue’ and a chromatic colour is ‘perceived colour possessing hue’. Furthermore, this contrasting effect is in certain situations enhanced as object
11 placed on a green background appear redder, on a red background appear greener, on a blue
background appear yellower, and on a yellow background appear bluer (Fairchild, 2005). A simple example of this is given by Regan et al. (2001), where red is the colour of ripe fruit can be detected from afar especially against green background such as green leaves.
The colour research has been mainly focused on the influence of red on a scale of decision making processes and in most other behavioural procedures. The scale relatively broad, from that women wearing red attract men in a bar to red resulting in avoidance behaviour. The research regarding associations or actions evoked by green, are heavily understudied. However, the most colour research uses green to contrast red. Further elaboration on the potential effects of green colour priming are not provided. The research conducted in this research regarding green should be evaluated as explorative.
The associations people have with red and green are different. Red is mostly are colour depicting danger or loss and green is mostly a depiction of gain and growth in nature(Moller, Elliot, & Maier, 2009). However, depending on the situation one is finding him- or herself in, colour
associations can differ. The colour red is a peculiar case within this respect as women wearing red lipstick at a bar attract men(Guéguen, 2012) but red is also an indication of aggression (Bruno, Martani, Corsini, & Oleari, 2013). There is a lot of research done on the colour red but the colour green is not that much investigated, the colour green is mainly used as contrast colour to red. Although there are studies that found that green associates with the term ‘go’ as this appears frequently in their lives as for example in traffic light (Elliot et al., 2007). For the colour green there are no more supplementary studies that link green to risk behaviour or performance. Even in the study of Kliger & Gilad (2012), when colour priming was used to see if this alters investment decisions, the potential effects of the colour red are elaborated in contrast to the effects of the colour green, which are very briefly or not discussed at all besides mentioning that it is a contrast colour of red. This raises the question if there is as much research conducts with respect to the colour green and its behavioural implications.
For the colour red, there is more research conducted and sometimes is delivers contradicting findings in for example performance behaviour where red worsens the performance (James & Domingos, 1953; Nakshian, 1964; Sinclair, Soldat, & Mark, 1998; Soldat, Sinclair, & Mark, 1997) and other find better performances(Hill & Barton, 2005). Red is also associated with mistakes, danger, and hazardous situations(Elliot, Maier, Binser, Friedman, & Pekrun, 2009). This association is assumed to be product of multiple sources one encounters in daily life. The most direct and specific interaction with red and failure start at a young age in the educational system, as mistakes and
12 failures are marked in red (Elliot & Maier, 2007). Even simple typo’s in word are underlined in red as indication that something was wrong, a mistake or a failure to type the word right. In everyday life, red is often used to convey negative information and this takes various forms in various situations such as sirens, traffic light, terrorist alerts, financial statements, stock markets, alarms, and warning signals. In other situations red conveys objectively danger in the form of an angry face, fire, and exposed blood(Moller et al., 2009).
The association of red with danger is strengthened over time as red appears in cases where negative consequences are salient such as the red depiction of a fire truck, red alarm lights, red traffic lights and the warning signs that are lined in red. Society uses red to warn for dangerous situations or failure and so on could be emerged from a biologically based indication that red is viewed as signal of danger(Elliot, Maier, Moller, Friedman, & Meinhardt, 2007). There is research that provides evidence that when facing a negative object, event, or possibility this evokes avoidance behaviour of this object, event, or possibility (Bargh & Chartrand, 1999; Cacioppo, Gardner, & Berntson, 1999). As the perception of red in achievement contexts indicates a possibility of failure it evokes a tendency not to take the risk (Elliot et al., 2007).
Although, there are examples that colour associations are situation depended. Such as red may enhance attractiveness of a potential mate regarding sexual encounters but has the opposite effect when evaluating a person’s competence. In the latter situation red is associated with failure/danger as in the former situation it is associated with sex/romance(Elliot, 2015).
Another perspective that is posited in the literature is the conceptual metaphor theory of colour (Meier & Robinson, 2005). This perspective evaluates the use of people’s speech when using colour to define abstract concepts in concrete terms grounded in experience to help them
understand and navigate the social world. Examples of such are ‘seeing red’ when describing anger as it entails that anger makes the face red. The same is applied to the concept of light and dark, light is metaphorically linked to good and dark to bad. ‘Seeing the light’ is a figure of speech to say when a situation is turning for the good and ‘in the dark’ implicates that the situation is not good (Lakoff & Johnson, 1999).
The results in the literature regarding associations with green are less in numbers as well as in clarity. In the research with word association by Moller et al. (2009), where words depicted in green, white, or red, there was just a single association observed for green. The association was that it was positively linked to success. Although, they mention that this could be due the culturally constrained factors of the U.S. undergraduate students. As in the U.S. green is shown when financial success is at place as well as in their currency (i.e., “greenbacks”). However, this associations could
13 be applied to other Western developed countries as these situations are also depicted in green in contrast to some Asian countries where financial success at the stock market is depicted in red. Other associations that are possible with the colour green are in the meaning of ‘go’ by traffic lights and it connection to growth in nature(Moller et al., 2009). Moller et al. (2009) acknowledge that hue priming and colour more generally is dramatically understudied but they provide evidence that hue is a nonlexical stimulus can transfer information subtly, quickly and potentially across barriers of age, language and may even species. Furthermore, they theoretically do not explore green associations in depth due to the lack of studies conducted in this area but they do, as many others, use green as contrast colour to red as this in general accepted by the colour appearance model of Fairchild (2005).
To sum up, red evokes avoidance behaviour and is linked to different kinds of danger and hazardous situations that this will probably leads to more risk averse behaviour when red is primed in the task. The opposite holds for green, although there is not much research conducted in the association between green and avoidance behaviour so it stressed that this is purely explorative. However, green’s general association with ‘go’ or approval it is expected that green colour priming will result in more risky behaviour. The colour grey is the control colour as no associations are made between grey and behaviour relevant for this study, according to well-established colour models (Fairchild, 2005; Fehrman & Fehrman, 2004).
2.2 Risk Behaviour Assessment.
In order to see if colour priming as an effect on risk behaviour the most ideal environment to test this is in a setting where we can isolate the priming treatment. From the risk measurements earlier discussed the revealed preferences method is the best suitable to answer the research question. This mostly takes the form of a laboratory experiment. In this environment we can filter out other factors that can affect risk behaviour. Furthermore, this provides us with the opportunity to see if colour priming has an isolated effect which can be utilized as a stepping stone for further research where other variables are added. In order to find the best suitable experimental method to measure the potential effect of colour priming on risk behaviour, the most frequently used risk elicitation tasks are evaluated.
2.2.1 Risk Elicitation Tasks
In a study by Crosetto & Filippin (2016), four different risk elicitation tasks are discussed and the differences are set out. Each of the risk elicitation tasks can be utilized to measure risk behaviour and many scholars have applied these. The tasks each have characteristics that makes them suitable for the purpose of this research. We now, by the use of the analysis provided by Crosetto & Filippin (2016), determine which of these fits the best by answering the research question of this study. The
14 four risk elicitation tasks that are discussed (Crosetto & Filippin, 2016):
- the multiple price list (Holt & Laury, 2002) (Henceforth, HL); - ordered lottery choice task (Eckel & Grossman, 2002, 2008) (EG); - Investment Game (Gneezy & Potters, 1997) (GP);
- Bomb Risk Elicitation Task ( Crosetto & Filippin, 2012) (BRET).
Of course, there are multiple other risk elicitation task but Crosetto & Filippin (2016) focussed on this four for several reasons2. The reasons they mention are that these are the most commonly used, are fast and easy to implement, and result in a lower relative cognitive load for the subjects. However, the last reason is arguable as it not easy to draw the line between low and high cognitive load. Crosetto & Filippin (2016) acknowledge this but they use the simple argumentation that a task involving a few clearly spelled out choices is less cognitive demanding than tasks involving multiple choices. For this reason, that these tasks are less cognitive demanding, their appraisal of these risk elicitation methods fits to this study. The subjects of this research are not familiar with an experimental procedure, so to keep attention optimal, the procedure should not be cognitive heavy. As the original goal was to investigate the intergenerational transmission of risk behaviour and if this was affected by colour priming. However, due to time constraints this was not feasible.
First, the HL multiple price list lottery. In this risk elicitation task, subjects are required to make a series of choices between lottery pairs. In general, Option A is safer than Option B. However, at some point of the series of choices, the expected return of Option B is higher than that of Option A. At this tipping point, risk neutral subjects should switch. Risk loving subjects switch before the tipping point and risk averse do not switch at the tipping point. Most commonly the HL has a set of 10 choices and the tipping point is at the fifth one. Consistency in this task is that each subject just changes one time in the lottery choice. If subjects change more than one time their lottery of choice this is deemed as inconsistent and is commonly observed. This has implications for the data as it can result in data losses or the need for a stochastic decision model. At the end of the experiment, one row is randomly chosen for payment and the lottery is played to determine the pay-off of the subject.
Second, the EG ordered lottery selection. In this task, subject are asked to choose one lottery out of an ordered set. In the EG task, subject can choose out of 5 lotteries which are characterized by
2 Crosetto & Fillipin (2016) mention a few other risk elicitation task that they not discuss among them are
these: random lottery pairs (Hey & Orme ,1994); the Becker-DeGroot-Marschalk mechanism, auctions, and trade-off method (Wakker & Deneffe, 1996).
15 a linear increasing expected value as well a standard deviation of this expected value. The variation of the expected value is obtained by differentiating outcomes of each lottery but keeping the probabilities of each outcome the same (50%). In the laboratory subject have to choose one lottery out the presented five. A risk neutral subject should choose the lottery with the highest expected value. In most cases of EG, the lottery with the highest expected value has also to most risk attached to it. This can result in that one can only divide the subjects into two groups; either risk averse or risk neutral (Crosetto & Filippin, 2016).
Third, GP investment game. GP constructed an experiment in which the choices are framed as an investment decision. At the start of the experiment subjects are endowed with an x amount of tokens/euros/dollars. Subsequently, they have to choose to invest it in either a safe account or a risky account. The risky account can yield 2,5 times the amount invested in it but with equal probability it can yield zero. The save account yields just one time its investment. As the expected return is higher in the risky account, a risk neutral subject should invest all of his/her endowment in the risky account. This can thus result in the same problem as was mentioned by the EG task, namely that one can only divide subjects into two groups; either risk neutral or risk averse.
Fourth, the BRET. Subjects face a 10 by 10 square in which each cell represents a box, thus a boxfield of 100 boxes. Subjects are informed that 99 boxes are empty while one contains a bomb. The bomb is programmed to explode at the end of the experiment, so after the choices of the subjects have been made. There are two distinct versions of the BRET, the dynamic and static version. In the dynamic version, the subject presses the button start on the screen and boxes begin to open. The subject can stop whenever he/she wants. The boxes open in preset order, starting at the top left corner of the boxfield moving towards the right hand side and continues at the left hand side. Furthermore, the dynamic version shows the probabilities transparent as the amount of boxes opened and the remaining boxes are displayed. In the static version, subjects fill in the amount of boxes they want to open. For both versions, each box has a value and the total value of the boxes opened is counted at the end of the experiment. The subjects get this pay-off at the end of the experiment unless they have also opened the box that contained the bomb. In the scenario that they opened the box which contained the bomb, their pay-off is zero.
2.2.2 Comparison
Each of the tasks has its strengths and weaknesses and we are evaluating them in order to see which task fits the best the purpose of the research. Besides only the technical implications of the tasks, the implementation of colour priming is taken into consideration. The comparison is led by the findings of the analysis by Crosetto & Filippin (2016).
16 The range of risk attitudes is defined as if the task can measure the three different sub-domains of risk preferences; risk-averse, risk-neutral and risk-seeking. Out of the four elicitation methods, two of them allow to estimate an almost complete range of preferences. This are the HL and the BRET. The GP and EG can only measure preferences in the risk-averse domain and to an extent in the risk neutral domain. As their range is going from very risk averse to risk neutral options to opt for.
Precision of the elicitation methods increases with the number of choices as this decreases the measurement error of the parameter estimated. Within this respect, EG task has only five options to choose from and HL classify their subjects in ten different categories. In this case, the precision of the task can be at stake. The GP and the BRET, in contrast, can estimate risk preferences almost continuously. Furthermore, Crosetto & Filippin (2016) did conduct some simulations and introduced noise.
Noise can be reflected as the level of observed inconsistencies. In their findings, they report that of all the subjects in HL, 17% switches more than once and that around 5% did always choose the safe lottery. In the BRET they observed one dominated choice but do not further elaborate on the issue. The other two tasks do not allow for the detection of inconsistent choices, so, therefore no exhaustive comparison can be made. Another approach of testing the noisiness of the task can be approximated by means of the structural Maximum Likelihood estimations. From this approach, only HL has significant more noise than BRET and between BRET, EG, and GP no significant differences are found. However, when the inconsistent choices form the HL task are filtered out, the differences become indistinguishable. In conclusion regarding noise, they do not find any significant differences among the four elicitation methods when controlled for inconsistent choices.
Some characteristics of the tasks can have adverse consequences. First, the availability of a riskless alternative. This is when a task has a focal safe point in their choice menu. This affect the behaviour of subjects. It could induce certainty effects. People tend to disproportionally choose for the safe option than for a slightly riskier one which can lead to a lower measurement of risk they are willing to take (Andreoni & Sprenger, 2011, 2012). Furthermore, the focal save point could act as a reference point against the potential outcomes, gains as well as losses, can be evaluated(Crosetto & Filippin, 2016). Among the four elicitation methods discussed, two include a clear safe option namely; the EG and GP. The HL methods does not has a clear safe option but the lower outcome of the safest lottery can act as a focal save point. This can represent the minimum amount that subjects can earn with probability one, although they have to opt for Option A(not the risky lottery). The BRET, however, has no focal save point.
17 For the cognitive load of an experimental procedure, Crosetto & Filippin (2016) compare the compound lotteries with the simple lotteries. The reason why they compare this is that subjects of experiments are not indifferent between the two (Kaivanto & Kroll, 2012). Subjects behave more risk averse when a compound lottery is presented. Furthermore, Kaivanto & Kroll (2012) suggest to use one-shot experiments rather than repeated tasks. However, for most scholars it is impractical to use pure one-shot experiments despite the implications, both theoretical as empirical, for the payment protocol (Harrison & Swarthout, 2014). For the purpose of this research, it is best suitable to do just one experimental procedure. Otherwise, the probability of subjects recognizing the colour priming as the interest of the study which can have adverse consequences as the ‘demand’ effect could affect the results.
Despite the cohesive analysis, Crosetto & Filippin (2016) cannot single out one best task. As one single all-purpose risk elicitation task might even well not exist. However, they formulated some recommendations for researches to choose the best elicitation task for their needs. To avoid adding noise to the task, it should be simple to understand especially when the numeracy of the subjects is an issue. From that point of view, HL could be troublesome as inconsistencies are common within this task. Although, when the inconsistent answer are excluded, HL does not show different levels of noise in comparison with the other tasks. Nevertheless, it results in data loss. Furthermore, when researching the whole spectrum of risk preferences from risk-averse to risk-seeking, Crosetto & Filippin (2016) recommend HL and BRET. These two have the biggest range in contrast to EG and GP, who have only the range from risk-averse to risk-neutral. Moreover, when a save option is not desirable for the scope of the research, EG and GP does not hold and the BRET and HL are more applicable. The BRET has a better understandability-precision trade-off than HL in that respect but HL provides more robust estimates.
For the purpose of this research, some characteristics are of interest. Both the colour red and green utilized to investigate if they affect risk behaviour. Red may induce more risk-averse behaviour and green more risk-seeking behaviour. Grey is the control colour, so subjects should behave risk neutral in when grey is depicted. Furthermore, as the subjects of this research are parents from a primary school, understandability is of interest. Not all parents are expected to have the numeracy level of undergraduate students, when assuming a normal distribution of the population. Within this respect, an one-shot experiment is preferred over a multiple choice alternative as they are less cognitive heavy. Weighting all the needs for the research, the BRET is the best fit. It does not contain a save option, it provides the whole spectrum of risk preferences to be investigated, and is not cognitive heavy as it is an one shot game. The HL would be second best, but the consideration of inconsistent choices by an less numerate subject pool can be high and inherent data loss is an
18 consequence when choosing HL (Crosetto & Filippin, 2016). The BRET is the best fit as risk elicitation task but can it also be altered to implement colour into the experimental design? Yes, is the answer. The colour of the boxfield, the main focus point of the experiment for the subjects, can be changed of colour. The default colour of the experimental design is grey, but this can easily be changed to either red or green. So, all theoretical and empirical considerations taken into account plus the availability to implement colour into the design, the BRET is the risk elicitation task that is utilized in this study.
2.4 Control questions
To control for potential effect that may cancel the colour priming effect out, these questions are asked in the experiment. In the study of Lejuez et al. (2003) the question concerning risky behaviour were simple yes and no questions and the risk taking questions were divergent in fields of health, financial and criminal activities. The questions they asked were all about if the participants engaged in these actions is the past twelve months. The ten questions they had were concerned about: smoking a cigarette (even one puff); alcohol consumption (even one drink); use of illegal drugs; gambled for real money; sex without a condom; carrying any weapon outside of their residence; had been in a physical fight; not wearing a seatbelt in a car; not wearing a helmet when riding a motorcycle of bicycle (even once). Not all of these subjects are addressed in the control questions of this research as some of the subject can of sensitive content for an ethnically diverse subject pool. These questions are asked after the experiment otherwise the demand effect could arise and subjects could already know the goal of the experiment before it started. Although this was a questionnaire concerning young adolescents, adults risk behaviour questionnaires ask around the same with regards to substance use (smoke and/or drugs), sexual behaviour (health), driving styles (health) and general risk behaviour (Bradley & Wildman, 2002). Additionally, other questions can be asked, such as; if participants did not stop at a traffic light when they were supposed to and if they drove more than 20 mph above the speed limit (Lejuez et al., 2003). However, the other side of the coin is that the overlap between stated and revealed preferences is considerably low (Crosetto & Filippin, 2016).
19
2.5 Hypotheses.
The hypotheses in are formulated in this subsection of the theoretical chapter. There are three hypotheses as the grey BRET is the control one and this study investigates if risk behaviour is significantly altered by the use of red and green. So, the hypotheses are regarding the differences in the amount of boxes opened between grey-red, grey-green, and red-green.
H1: When the boxfield is depicted in red, subjects behave more risk averse than in the control task. Subjects behave more risk averse as the boxfield is depicted in red than the control group does with the depiction of the boxfield in grey. So, subjects interacting with a red boxfield are expected to significantly take less risk than the subjects interacting with a grey boxfield.
H2: When the boxfield is depicted in green, subjects behave more risk prone than in the control task. Subjects behave more risk prone as the boxfield is depicted in green than the control group does with the depiction of the boxfield in grey. So, subjects interacting with a green boxfield are expected to significantly take more risk than the subjects interacting with a grey boxfield. H3: Subject behave more risk averse when the boxfield is depicted in red than in green.
Subjects behave more risk averse as the boxfield is depicted in red than the other treatment group with the depiction of the boxfield in green. So, subjects interacting with a red boxfield are expected to significantly take less risk than the subjects interacting with a green boxfield.
20
3. Experimental design
The method used in this study is an experimental and this section of the paper sheds light on the set up of the experiment and how it connects colour priming and risk behaviour. The advantages and disadvantages are also discussed as well as the internal and external validity. As previous research mainly had the colour priming aspect on the background, the difference in this study is that the main part is depicted in colour(Kliger & Gilad, 2012). This may affect risk behaviour in a more substantial way than previous research. The experiment of use is the Bomb Risk Elicitation Task (BRET) by Crosetto & Filippin (2013) which is a well-known risk experiment. The experiment is facilitated to the subjects by using the Qualtrics software. By using the static version of the BRET by Crosetto & Filippin (2015) the amount of risk a person is willing to take can be better measured than in de dynamic version as the bomb can explode before the person was willing to stop. This leads to information loss so that is why the static version is preferred. However, the connotation of the bomb is changed as this connotation could evoke reactions that drove the attention away from the task. As the population of this experiment is diverse in ethnicity, some can may find this offensive. So, in this case connotation of the ‘bomb’ is replaced with a ‘waterpump’ and the tokens are replaced with ‘paper tokens’. When a subject activates the waterpump all opened boxes with a paper token in it become worthless. Thus the general message is the same and does not change the experimental task itself. After the experimental tasks, the control questions are asked. The instructions as well as the general layout of the experiment are adapted from the BRET study by Crosetto & Filippin (2013).
The design of the experiment is a within-subject design as the subject answers the control task first and after that the treatment task is asked. All participant conduct the control task in grey and half of them conducts the red treatment task and half of them conducts the green treatment task. So, the probability for a subject to be either in the red or green treatment is fifty percent. This provides the opportunity to see if colour priming has indeed a direct effect on risk behaviour when setting it out against the control task. Furthermore, the scores of the treatment groups can be compared to one another to see if it makes any difference if the same task is differentiated only in chromatic colours.
The experimental set up was in a classroom on a primary school located in Eindhoven, pictures of the set up are shown in Appendix 1. It was on the opening day of the new building of the school and the subjects were the parents of the children that went to that school. On beforehand it was expected that around 80 till 100 parents would come to the opening day of the new building. However, not that many showed up and some data had to be dropped due not completing the task or filling in 1 or 100. The tables in the classroom were adjusted in a test setting so that parent could
21 not collaborate and provided oversight. The school provided us with chromebooks for the facilitation of the qualtrics software what was used in this experiment.
Before the experiment could take place, the parents were updated on the policy of the Radboud University and the codes of the law regarding privacy and the use of the data. This took more time than anticipated before and by some parents the attention drifted away. After the
introduction and information of the policy and codes of the law, the experiment was introduced. The experiment was introduced by explaining the experimental procedure to all participants. After that, some test questions were asked to see if participant get how the experiment works. The answers the participants gave on the test questions were provided by feedback depending on the answer itself. If a participant answered the question wrong, the feedback showed why it was wrong in words as well as visual. This visual display was the boxfield itself, the amount of boxes opened and the location of the waterpump. When the explanation and test questions were finished and no further questions were asked, the experiment could start. All participants started at the same time with the control task which was followed up by the treatment task which was either depicted in red or green. After the control and treatment tasks, control question were asked regarding risk behaviour in real life. Which task would be paid out to the participant was randomly chosen by the computer and announced at the end of the experiment. After the control questions because the announcement could otherwise influence the answers the participants would gave.
3.1 Colour use.
The use of colour in the experiment regarding risk behaviour is the main goal of the study and therefore this experiment is used to determine if colour priming significantly affects risk behaviour. From well-established colour models, green and red are counter colours and from other psychology and economic literature green and red recall different associations like the gain and loss domain. For the control task grey is chosen due to the lack of associations it recalls by participants according to the well-established colour models (Fairchild, 2005; Fehrman & Fehrman, 2004). So the use of colour is implemented at the main object of the task, the boxfield. It is thus depicted in grey in the control question and then in either red or green. Grey as control task and green and red are the colours of interest to see if they affect risk behaviour in a significantly different way than the control task.
3.2 Experiment pay-off.
The experimental pay-off function is the same as in the paper by Crosetto & Filippin (2013). The instructions of the experiment are reformulated to make it understandable for the subject pool.
22 The instructions can be found in Appendix 2. The pay-offs in this experiments are paper tokens, which are converted at the end of the experiment to Euro. The money is not paid out immediately but, if the participant has earnings, it will be deducted from their school trip fee of their children. In this way, the monetary incentive is still in place and the requirements of salience, monotonicity and dominance are preserved. The experiment instructions are in Appendix 2.
After the instructions, some practice questions are asked to see if the subjects understand the experimental procedure. Two different hypothetical scenarios are asked where no tokens can be earned or loosed. The subject can collect boxes but if the waterpump is activated, the earned tokens become useless. So, the collect boxes Bc cannot be equal or higher than the number in which the waterpump Bw is. So the payoff function is as follows:
0 𝑖𝑓 𝐵𝑐 ≥ 𝐵𝑤 𝑜𝑟 𝐵𝑐 𝑖𝑓 𝐵𝑐 < 𝐵𝑤
The waterpump is collected in two of the hypothetical scenarios and in one the waterpump is not collected. The first question is; Suppose you collected 34 boxes. The waterpump was in box 58. Did you earn tokens? Yes or No. If the subject understood the instructions well, they would answer yes as 34 is less than 54 (34<54), so the waterpump was not activated. The second question is; Suppose you collected 76 boxes. The waterpump was in box 44. Did you earn tokens? Yes or No. In this hypothetical situation, the waterpump was in box 44 and the subject collected 76 boxes and thus activated the waterpump and lost the tokens earned. So, the subject had to answer ‘No’ to the question as 74 is more than 44 (74>44).
When subject do not answer this question correctly this can indicate that they do not understand the experiment. However, due to the practice questions this can be minimized. These practice question are followed by visual appearance of the situation sketched in the question, to make it easier for subjects to understand the mechanism. So, if they do not correctly answer the first question, see the visual depiction of the situation and receive feedback. Everyone receives feedback on their given answer in the practise question. The boxfield is displayed, with locations of the waterpump and how many boxes were collected in the practice question.
3.3 Empirical Method.
The empirical strategy of the analysis the data is set out in this part in order to find evidence that confirms the hypotheses stated in at the end of the second chapter. The main difference between
23 the subjects is the colour of the boxfield and it is expected that this affects risk behaviour. Due to the within-subject design, every participant has the control task and a treatment task. Half of the
participants have the red and the other half have the green treatment task. Red and green evoke different associations regarding risk behaviour as discussed in the previous chapter. The grey group acts as a control group, as their results are not affected at all by any colour use as grey is a neutral colour. This provides the opportunity to see if colour has indeed an effect on financial risk behaviour. Furthermore, to see if the amount of collected boxes differ between the control and treatments questions and between de treatments questions, a repeated measures ANOVA is utilized. This analysis technique is commonly used for a within subject design and is used to test the first two hypotheses (Kherad-Pajouh & Renaud, 2014). The baseline treatment in this study is the grey colour task as no associations are evoked by the use of grey. For the third hypothesis, the Mann-Withney U test is applied. This method was used in the original paper of Crosetto and Filippin (2013) to
investigate if the different treatments differ by the amount of the collected boxes. The different treatments in their research were regarding the static and dynamic version as well as the low and high stakes versions they let the subjects conduct, the baseline and fast treatment, and mixed amount of boxes in boxfields. This analysis is useful in the sense that the subjects get different treatments by colour in either green or red. Pairwise correlations are applied to see if the results found have external validity. Crosetto & Filippin (2016) discussed the overlap between the revealed preferences and stated preferences. The overlap between the four risk elicitation tasks and
frequently used questionnaire set was around 20 to 30 percent, which indicates that there is still a discrepancy between the revealed preferences and stated preferences regarding risk behaviour.
Besides the aforementioned colour priming, the control questions are taken into account when conducting the empirical analysis. These can provide additional information if other factors also influence risk behaviour or that these factors are solely responsible for the effects found in the analysis. Moreover, it can also be the case that these factors enhance the influence of colour priming instead of cancelling them out. Although, it can also be the other way around. That colour priming enhances the effect of the factors that influence risk behaviour in the first place. Or it can be the most peculiar case that risk taking in situations other than in the financial domain do behave more risk averse in financial situations.
24
3.3.1 Methodology.
3.3.1.1Dependent Variables.
The dependent variables are the amount of boxes collected in the control task and the amount of boxes opened in the treatment tasks. This is the measure of how much risk each participant was willing to take to maximize their potential pay-off as they can always active the waterpump.
The amount of boxes collected in the control task can range from 1 to 100. When collecting 100 boxes, the waterpump is certainly activated, resulting in all token becoming worthless. For the treatment tasks, the range is 1 to 25. Collecting 25 boxes activates the waterpump without a doubt and the tokens become worthless. The change in the size of the boxfield is to minimize the potential demand effect. In the situation that the boxfield remained the same size and only the colour
changes, subjects could notice and behave accordingly. These three variables are the variables we need to use to answer the research question.
The analysis is aimed at investigating if participants opened less boxes in the red treatment task than in the control task, and, if participants opened more boxes in the green treatment task than in the control task, and, if participants opened less boxes in the red treatment case than in the green treatment task.
3.3.1.2 Independent variables.
A dummy variable is made for the colour treatment in order to compare differences among the amount of boxes opened between the treatment groups and the control groups. So, one dummy were the participants who conducted the red treatment task were given a value of 1 and the
participants who conducted the green treatment task were given a value of 0. For the use of the Mann-Withney U test, the variable Treatment has either the text Red or Green showing which treatment the participants had.
For the answers of the control question take were asked after taking the experiment, an index is constructed. The control questions were concerning their behaviour on the road by either using a bike or a car. The question that were asked are displayed in Appendix 3. The answer of these question could have a value ranging from 0 to 5. For these values, the value of zero says that the participant did not use either a car or a bike for transport thus could not answer the question as such. For the values 1 to 5, the higher the value the more risk taking the participant was assumed. Each of the tree questions has an equal weighting regarding their influence on the index. The index is constructed as follows: the values of the questions are added and divided by the total number of
25 question giving an average indication of risk behaviour in outside of the financial realm. The average is corrected for the participants that answered a question that they never use the car or bike thus giving a zero value to the answer of the question. When answered zero, no indication can be drawn from their answer. This index is called the ‘risk index’. Besides one total risk index of all participants, two additional indices are constructed, each for one treatment group. Should these significantly differ in average risk taking, the amount of boxes could be affected in the way that one group already is more risk prone or averse. Which can subsequently implicate the findings of this study.
3.4 Method.
Due to the within-subject design for the control and treatment groups, we use a commonly utilized test to see if the scores of the two tasks differ. The commonly utilized test is a repeated measures ANOVA (Kherad-Pajouh & Renaud, 2014). Since the sample size in this study is small, a repeated measures ANOVA is specifically designed for a within subjects design. A subject pool of 21 subjects has in a within subject design 42 observations. Crosetto & Filippin (2013) used such a test to investigate differences between the groups they had, namely the Mann-Withney U test. The
assumption underlying this test is that the sample is randomly selected from the population, the observations are independent and the measurement should be of at least and ordinal scale. For the investigation of the results found in the study have external validity, the answers on the control questions are used to construct an index. By the use of a pearson correlation, we can investigate if the scores of both groups as well as the group as a whole has external validity. However, Crosetto & Filippin (2016) have investigated the correlation between the risk elicitation tasks and two risk questions indexes and the correlation over the whole spectrum was low. This suggests that there is still gap to be bridged between the revealed and stated preferences within this respect.
26
4. Results
In this section, the data is tested to see if there is statistical difference in the amount of risk the participants take in the control task and treatment tasks. Let us first start with investigating the descriptive statistics to see if any abnormalities can be seen. The descriptive statistics are shown below. When comparing the mean of the control task with the mean found in the original paper of the BRET by Crosetto & Filippin (2012), we notice that our mean is slightly higher than they observed namely 40 by 45 roughly speaking. Furthermore, the maximum amount of boxes a subject could open was 100, thus certainly activating the waterpump. This occurred one time, for further analysis purposes we analyse the data with and without this outlier to see if it causes any changes. This also applies to amount of boxes opened in the treatment task of this subject. The subject may did not yet understand the task at hand despite the two practice questions. Moreover, the range of the red and green treatment task differs slightly, with red having a broader range than green as well as a higher standard deviation. The mean of the green treatment task is higher than the mean of the red treatment task. This could imply that a difference is present due to colour priming but that is discussed further in this section.
Table 1: Descriptive Statistics
Variable Obs Mean Std.Dev. Min Max
Control 42 44.548 18.749 12 100
Red 21 10.048 4.79 4 20
Green 21 11.905 3.3 6 18
Total Risk Index 42 1.429 .595 0 2.667
Red Risk Index 21 1.492 .574 .333 2.333
Green Risk Index 21 1.365 .623 0 2.667
. In Figure 1, the normality is displayed. The graph of the red scores seems slightly skewed. However, if this is not significantly different from the normality itself, is should not have any complications to pursue the analysis forward. We tested this and observed that it not significantly deviates so we can proceed with the analysis. Furthermore, in the control group one subject opened 100 boxes and thus activated the waterpump without a doubt. This is probably because the subject did not yet understand the task at hand well enough. In the treatment task this subject however, gave a reasonable answer. Therefore, the first observation in the control task is not included in the analysis. For the answer of the subject in the treatment task, the analysis is conducted with and without to see if this has potential effects on the results found. Before we go to analyse the data in depth, we first need to check if the data is normally distributed and not skewed. Figures 1 and 2 display the distribution. Figure 1 displays the distribution of the control task and figure 2 displays the distribution of the red and green treatment tasks. Besides the visual display of the normality, it was tested if the data significantly differs from the normal distribution but this was not the case. So for
27 the control scores, we can pursue with the analysis without any restrictions regarding normality. The same applies for the skewness of the control scores3.
Table 1: Normality Control
For the distribution of the treatment tasks, the same procedure applies. The distribution is displayed in figure 2. The green treatment scores are well normally distributed. For the scores of the red treatment tasks it seems this is not case. However, when tested is this deviates too much as well if it skewed, the test did not indicate this was the case. So, we can proceed with the tests that were anticipated in the previous section4.
3 Tests are in the Appendix 4. 4 Tests are in the Appendix 5.
28
Table 2: Normality treamten tasks
We now start with the analysis for hypothesis one, which is a repeated measures ANOVA. The first hypothesis states that the subjects open less boxes in the red treatment task than in the control task. However, the sign in expected, the repeated measures ANOVA does show that this effect is not significant indicating that colour priming has no significant effect when red is compared to grey5. So, the risk behaviour of subjects is not significantly affected by the presence of red colour priming.
Second, we conduct the repeated measures ANOVA for the second hypothesis. The second hypothesis states that the subjects open more boxes in the green treatment than in the control task. In line with the first hypothesis, the sign is as expected but the repeated measures ANOVA shows that the effect is not significant. This indicates that colour priming has no effect when green is compared to grey6. Thus, the risk behaviour of subjects is not significantly affected by the presence of green colour priming.
Third, The Mann-Withney U test is utilized to test the third hypothesis. The third hypothesis states that subjects in the red treatment group open less boxes than the subjects in the green treatment group. The difference with the first two hypothesis is that the analysis is not longer a within-subject design as these scores are from different subjects resulting in a between-subject design. The results of the Mann-Withney U test shows a significant difference between the red and green treatment group at the 10% confidence interval (0.0808)7. The difference between red and
5 For output, see Appendix 6. 6 For output, see Appendix 7. 7 For output, see Appendix 8.
