The influence of
happiness on
financial risk tolerance
The influence of happiness on
financial risk tolerance
Master thesis Finance
By: Juliët Nanninga
University of Groningen
Faculty of Economics and Business
Abstract
Most research on financial risk tolerance ignores the potential influence of feelings. Therefore, this thesis investigates the influence of happiness on financial risk tolerance. It uses data from the DNB Household Survey to create a sample with 20,376 person/year observations over the period 1995-2013. The results show that respondents who report the highest level of happiness, display the lowest level of risk tolerance (or: highest level of risk aversion). These findings are consistent for two different measures of risk tolerance.
Table of contents
1 Introduction ...1
2 Theoretical background ...3
2.1 Literature review ...3
2.2 Hypothesis development ...9
3 Data and methodology ... 10
1
1 Introduction
Traditional finance theory describes people in a way that most closely resembles the character of Mr. Spock from the popular television and movie series Star Trek. Mr. Spock is half Vulcan, a species that suppresses emotion and prizes logic. He is a rational thinker who thoroughly considers every piece of information. In line with this, traditional finance theory assumes that people incorporate information into their decision processes using the rules of probability and statistics with calculated, unemotional logic (Ackert, Church and Deaves, 2003). People make rational decisions to maximize their wealth, or expected utility, in the face of risk and uncertainty (Nofsinger, 2013). The utility framework treats emotions as a factor that can be neglected (Guven, 2012; Grable and Roszkowski, 2008). This assumption about human behavior is unrealistic. Our world is not inhabited by Vulcans. Contrary to Vulcans, humans are emotional beings, who suffer from cognitive biases in decision-making. Our decisions are often driven by emotions. We let our mood influence all kinds of daily decisions, ranging from the type of clothes we wear, to the type of food we eat (Grable and Roszkowski, 2008). Therefore, it is not surprising that the traditional economic utility approach fails to adequately explain many financial attitudes and behaviors, such as the shifting of risk preferences when questions with similar payoffs are framed differently (Grable and Roszkowski, 2008).
So far, research in behavioral finance has primarily focused on cognitive biases in financial decision-making, paying less attention to the role of emotions (Lerner, Small and Loewenstein, 2004; Ackert et al., 2003). Examining cognitive aspects of financial behavior in isolation from emotions is troublesome and misleading (Ackert et al., 2003). Recent neurological and psychological studies provide strong support for the idea that emotions play a key role in decision-making (e.g. Loewenstein, Weber, Hsee and Welch, 2001; Forgas, 1995; Schwarz, 2000; Damasio, 1994). Emotions interact with the cognitive evaluation process that eventually leads to a decision (Nofsinger, 2013). The influence of emotions is the strongest, when the decision is complex and surrounded by uncertainty (Forgas, 1995).
Financial decisions are complex and include risk and uncertainty (Nofsinger, 2013). It seems likely that they are influenced by emotions. However, the influence of emotions on decision-making has received little attention within the finance literature (Ackert et al., 2003; Lerner et al., 2004). The studies that do look at emotions use macro-economic data (e.g. aggregate stock market outcomes) and focus on professional investors (e.g. Kamstra, Kramer and Levi, 2003; Edmans, Garcia and Norli, 2007; Hirshleifer and Shumway, 2003). These studies generally assume that some environmental factor (e.g. sunshine) is responsible for generating mood changes in a large fraction of the investor population, which in turn influence their financial decisions (Lepori, 2010). While these studies indicate that moods can influence financial decisions, they provide little insight at the level of the individual. Research that looks at the individual level is mostly found in the psychology literature, and does not focus on financial decisions (Lerner et al., 2004). This thesis addresses the gap in existing literature by investigating, at the individual level, the influence of happiness on financial risk tolerance.
2 Young, single, well-educated men with a good income and high financial knowledge are deemed to be the most risk tolerant (Grable and Roszkowski, 2008). The influence of mood is usually ignored. Psychology literature does acknowledge a relationship between mood and risk tolerance. However, there is no consensus about how the relationship exactly works. There are two competing theories: the affect infusion model (AIM) and the mood maintenance hypothesis (MMH). The AIM claims that a good mood increases risk tolerance (Forgas, 1995). According to the AIM, people in a good mood are more likely to access thoughts prone to positive aspects of risky situations. This leads to a more favorable assessment of the situation (Chou, Lee and Ho, 2007). On the contrary, the MMH argues that a good mood lowers risk tolerance (Isen and Patrick, 1983). According to the MMH, people in a good mood are less willing to take risks, because they want to maintain their current emotional state. Taking risks increases the potential for substantial losses, which are detrimental to a good mood (Chou et al., 2007).
The aim of this thesis is to provide more insight into the influence of a good mood (happiness) on an individual’s financial risk tolerance. Are happy people more risk tolerant? To answer this question, this thesis uses data from the DNB Household Survey. Each year, about 2,000 Dutch households fill in this questionnaire about their personal and financial situation. Data is available for the period 1993-2013. This survey lends itself well to investigate the topic, since it combines economic and psychological concepts. This thesis distinguishes itself from existing finance literature by focusing on households instead of investors, and using micro-economic data. It distinguishes itself from existing psychology literature by focusing on financial decisions and using large-scale survey data. Most psychological studies have an experimental design, in which mood states are induced and students are used as participants. Apart from academic relevance, this thesis also has practical relevance. More clarity on the influence of mood on financial risk tolerance can improve consumer decisions related to investments and allocation of assets (Grable and Roszkowski, 2008).
3
2 Theoretical background
2.1 Literature review
The terms mood and emotion are often used interchangeably, as they are both affective states (Waters, 2008; Forgas, 1995). However, they are in fact theoretically distinct constructs. Moods are low-intensive, diffuse, and relatively enduring affective states that often arise for no particularly salient reason (Treffers, Koellinger and Picot, 2012). Emotions are more intense and short-lived affective states that generally have a definite cause and clear cognitive content (Treffers et al., 2012). Thus, moods and emotions differ in intensity, duration and origin: emotions are feelings that arise in response to specific stimuli, whereas moods are free-floating feelings that need not be linked to anything specific (Grable and Roszkowski, 2008). The distinction between moods and emotions is more theoretical than empirical. Research practice often uses identical methods to induce moods and emotions (Treffers et al., 2012).
Affect is used as an umbrella term to refer to both mood and emotion. Affect can be defined as the specific quality of goodness or badness 1) experienced as a feeling state (with or without consciousness) and 2) demarcating a positive or negative quality of a stimulus (Slovic, Finucane, Peters and MacGregor, 2004). At the most general level, affective states can be categorized into positive (pleasant) and negative (unpleasant) feelings (Grable and Roszkowski, 2008). In the context of decision-making there is also a difference between integral and incidental affect (Waters, 2008). Integral affect includes the influence of moods and emotions that are normatively relevant to the present decision (Han, Lerner and Keltner, 2007). For example, the anticipated regret when evaluating a gamble influences how much one is willing to gamble (Han et al., 2007). Incidental affect includes the influence of moods and emotions that should be normatively irrelevant to the present judgments and choices (Han et al., 2007). For example, Hirshleifer and Shumway (2003) show that mood-states induced by weather conditions can influence stock market outcomes. This distinction shows that emotions and moods can inform decisions, even when their cause is unrelated to the decision being made (Lucey and Dowling, 2005; Nofsinger, 2013).
Affect and decision-making
4 decision-making motives (Raghunathan and Pham, 1999). Feelings can bias the value attached to certain outcomes (Seo and Barrett, 2007). Research shows that intense negative emotions make people favor short-term enhancements, regardless of possibly negative long-term consequences (Seo and Barrett, 2007).
Although emotional influences on decision-making are typically characterized as irrational, recent research suggests that emotions and rational decision-making are complementary (Ackert et al., 2003; Damasio, 1994). There is evidence that feelings improve decision-making performance by facilitating and even enabling decision-making processes (Seo and Barrett, 2007). Again, characters from Star Trek provide a good analogy. The Vulcan Mr. Spock works together with the human Captain Kirk. While Kirk is emotional, he is portrayed as a good decision-maker. Though Spock fully analyses each situation, he gets too caught up in the details. Emotion allows Kirk to focus and it enhances his ability to make critical decisions.
Damasio (1994) studied brain-damaged patient who retained their cognitive abilities, but had impaired emotional responses. As a result of frontal brain lobe damage, the patients were emotionally flat. However, their knowledge, attention, memory, language, and abstract problem solving was unaffected. Damasio (1994) showed that these individuals had difficulty making decisions and were unable to plan for the future or choose a course of action. Damasio’s work illustrates the striking effect that a lack of emotion has on decision-making.
Feelings drive conscious attention and allocation of working memory. Both are necessary for the extensive cognitive processes involved in decision-making (Damasio, 1994). Feelings facilitate in selecting and prioritizing choices relevant to situational requirements (Seo and Barrett, 2007). Potentially infinite factors and options surround every decision, each with conflicting advantages and disadvantages (Seo and Barrett, 2007). This can make it extremely difficult to make an optimal decision within a given time frame. An individual might become overwhelmed by the possibilities (Ackert et al., 2003). Feelings can efficiently and effortlessly guide our deliberations of the relative desirability of alternative options (Finucane and Holup, 2006). Feelings invoke distinguishable frames of mind, which help to selectively attend to and efficiently prioritize cues (Seo and Barrett, 2007; Raghunathan and Pham, 1999). Feelings help to anticipate the consequences of various actions, simplify scenarios, and resolve ambiguity (Finucane and Holup, 2006). In other words, feelings provide a coping mechanism that allows people to focus without being caught up in the details (Ackert et al., 2003). As mentioned earlier, feelings influence how people process information (Schwarz, 2000). This enhances decision-making effectiveness in particular contexts (Seo and Barrett, 2007). For example, people in pleasant affective states tend to categorize stimuli in a broader, more inclusive, and more flexible fashion. This improves creativity and performance on complex tasks (Seo and Barrett, 2007; Ackert et al., 2003).
Affect and risk
5 The past several decades, researchers have begun to examine the influence of affect on the way people perceive risk and the way they make risky decisions (Grable and Roszkowski, 2008). These researchers acknowledge that responses to risky situations and circumstances are a result of both analytical (cognitive) and affective influences (Grable and Roszkowski, 2008). A first advancement on the traditional perspective has been to include the impact of anticipated emotions (Lucey and Dowling, 2005). Anticipated emotions are emotions the decision-maker expects to experience, given a certain outcome. Examples are regret and disappointment. More recently researchers started to include the influence of emotions experienced at the time a decision is made on the decision-maker (Lucey and Dowling, 2005). New models of how people judge risky options emphasize the important role of affect (e.g. Forgas, 1995; Loewenstein et al., 2001; Slovic et al., 2004). They see affect as essential for navigating through a complex and uncertain environment (Finucane and Holup, 2006; Forgas, 1995).
Loewenstein et al. (2001) propose the risk-as-feelings-model. This model postulates that individuals evaluate risky situations using both cognitive and affective processes. These processes are interrelated: cognitive appraisals give rise to emotions and emotions influence appraisals. However, people’s emotional reactions to risk depend on a variety of factors that influence cognitive evaluations of risk only weakly or not at all. Cognitive evaluations of risk are based on probabilities and outcome valences. Outcome valences include anticipated emotions such as regret. Feelings about the risk (or: anticipatory emotions) are influenced by factors such as vividness, immediacy and background mood. Emotional reactions are also sensitive to probabilities and outcome valences. However, the functional relationships are quite different. As a result, emotional reactions to risks often diverge from cognitive assessments of those risks. When such divergence occurs, emotional reactions often dominate the decision process and drive behavior (Loewenstein et al., 2001).
Closely related to the risk-as-feelings model is the affect heuristic (Finucane, Alhakami, Slovic and Johnson, 2000; Slovic et al., 2004). Affect serves as a cue for many important judgments. Representations of objects and events in people’s minds are tagged to varying degrees with affect. These are global evaluative feelings of liking and disliking. People consult or refer to an affective pool, containing all the positive and negative tags associated with the representations, when they are judging risk. Pleasant (unpleasant) feelings motivate actions and thoughts to extend (avoid) the feelings. In other words, people’s decision-making is guided by the images and associated feelings that are induced by the decision-making process (Lucey and Dowling, 2005). Using affective impressions to make decisions can be easier and more efficient than weighing the pros, cons and outcome probabilities. Especially so, when the decision is complex or mental resources are limited. Because affect is used as a mental shortcut, affect can be classified as a heuristic (Finucane et al., 2000).
Positive affect and risk tolerance
6 frequency of risks. In other words, positive affect leads to an overestimation of the likelihood of positive events, whereas it leads to an underestimation of the likelihood of negative events (Johnson and Tversky, 1983).
The evidence is much more controversial when one considers studies in which, rather than having to estimate event probabilities, people are asked to make actual or hypothetical decisions under risk (Lepori, 2010). In other words: studies that look at risk tolerance, instead of perception of risk. Within the psychology literature, two opposing theories have emerged to explain the relationship between a good mood and risk tolerance: the affect infusion model (AIM) and the mood maintenance hypothesis (MMH). The AIM states that positive moods increase risk tolerance (Forgas, 1995), whereas the MMH suggests that positive moods lead to risk-averse behavior (Isen and Patrick, 1983).
The AIM suggests that a good mood increases risk tolerance (Forgas, 1995). This is caused by biases in cognitive processing and selective information retrieval (Chou et al., 2007). People in a good mood tend to focus on positive cues in the environment. Also, positive moods cue positive memories. Memories that are associatively linked to the current mood are more likely to be recalled and used. People in a good mood are more likely to access thoughts prone to positive aspects of risky situations (Chou et al., 2007). This leads to a more favorable assessment of the situation. Additionally, people in a good mood may rely more on heuristic information processing (Schwarz, 2000). As mentioned earlier, this information processing strategy is characterized by top-down processing, with high reliance on pre-existing knowledge structures and relatively little attention to the details at hand. Therefore, individuals in a good mood may be less aware of potential negative consequences of their decisions and a lack of careful and rational thought may intensify their risk-prone responses (Lepori, 2010). On the contrary, people in a bad mood assess risky situations more negatively. They also use a more systematic processing strategy.
According to the MMH, a good mood induces risk-averse behavior (Isen and Patrick, 1983). This hypothesis explains the effect of mood on risk tolerance through an innate desire to maintain a positive affective state and to mitigate a negative mood (Chou et al., 2007). People in a good mood want to maintain this positive affective state. Consequently, they are unwilling to take risks. Taking risks increases the potential for substantial losses, which are detrimental to their good mood (Chou et al., 2007). There are studies that document that a good mood decreases risk tolerance, even when this mood leads to an overestimation of the probabilities of winning (e.g. Nygren, Isen, Taylor and Dulin, 1996). On the contrary, people in a bad mood do take high risks, as the potential gains may uplift their mood (Chou et al., 2007). Alternatively, their risky choices may be the consequence of a state of depletion, resulting from engagement in active mood regulation attempts (Bruyneel, Dewitte, Franses and Dekimpe, 2009).
7 taking. When high-risk bets are considered, individuals in a good mood tend to be more risk averse. Thus a good mood fosters different risk preferences, depending on the riskiness of the bet. According to Forgas (1995), the AIM is likely to dominate in contexts that require substantial information processing and an accurate evaluation of complex issues, whereas the MMH may dominate when people have a strong impetus to attain a particular outcome. Other authors suggest that the MMH is likely to dominate in decision frames where possible losses are real and salient, stakes are high and actual large losses may occur (Isen and Patrick, 1983; Lepori, 2010). A contingent theory about mood and risk-taking, considering the magnitude of forces at play in a particular context, may reconcile the two conflicting predictions (Au, Chan, Wang and Vertinsky, 2003).
Alternatively, the way affective states are measured may cause the contradicting results (Treffers et al., 2012). Most of the studies that address mood and risk tolerance use a valence-based approach. Valence describes the extent to which affective states involve pleasant or unpleasant experiences (Treffers et al., 2012). The valence-based approach fails to identify whether different emotions of the same valence will have a different influence on decision-making. For example, recent research shows that while fear and anger are both negative affective states, they have an opposite effect on risk perception (Lerner and Keltner, 2001). Fear seems to cause pessimistic risk assessments and leads to risk aversion. Angry individuals seem to evaluate risk more optimistically and therefore become more risk tolerant (Lerner and Keltner, 2001). Raghunathan and Pham (1999) provide another example with sadness and anxiety. These authors find that sadness results in higher risk tolerance, while anxiety results in higher risk aversion. These examples make clear that a more detailed approach is needed. Different affective states of the same valence can have a different impact on risk tolerance. Recently, researchers have begun to use the appraisal tendency approach (Lerner and Keltner, 2001). This approach predicts the effects of specific emotions, differentiated on the basis of their core cognitive appraisal dimensions such as controllability, certainty, novelty, pleasantness, or arousal-level.
Affect and finance
Most of the research on mood and risk tolerance is found in the psychology literature. These studies often have an experimental design in which mood states are induced, for example by letting participants watch sad or happy movie clips (e.g. Chou et al., 2007). Manipulating moods has the major advantage that it allows inference of cause-effect relationships. However, the induction procedures are rarely precise, and often give rise to concomitant changes in several emotional states. Additionally, the mood induction process may create some transient stress effects (Hockey, Maule, Clough and Bdzola, 2000). Researchers often use students as participants for their experiments. This limited sample may not be representative for the ‘real world’ (Lepori, 2010). Risk tolerance is often measured through lottery-type questions or overly simplified gambling tasks (Lepori, 2010). They provide an effective way of defining rational behavior, but may have only limited relevance to everyday choices, which normally have to be made in the face of uncertainty and ambiguity (Hockey et al., 2000). This is especially so for financial decisions, which are complex and include risk and uncertainty (Nofsinger, 2013). To put it differently, psychology studies do not look at affect and risk tolerance in the context of financial decision-making (Lerner et al., 2004).
8 factors as mood proxies, ranging from weather to hours of daylight and outcomes of sporting events. Kamstra et al. (2003) look at the effect of seasonal affective disorder (SAD) on equity returns. The authors use hours of daylight as an indicator of SAD and find a significant effect of SAD on the seasonal cycle of stock market returns. Edmans et al. (2007) find negative stock market reactions after losses in popular sporting events. Hirshleifer and Shumway (2003) find a significant positive correlation between sunshine and stock market returns. However, Kliger and Levy (2003) find a negative effect of good weather on risk attitudes in capital markets. These two studies show that, similar to the psychology literature, there are conflicting findings.
Overall, these studies show that moods can influence financial decisions. The findings of these studies challenge the assumptions that individual risk preferences are stable over time and across events with similar prospective outcomes (Treffers et al., 2012). However, the interpretations of their results remain vague with respect to which specific affective states are involved in the impact that is observed (Treffers et al., 2012). For example, a loss in a popular sport event can induce different emotions. Some people get angry when their favorite club loses, others sad. Similarly, weather can affect individual’s mood differently. Additionally, these articles look at market level data, not at individuals. Only few studies address mood and financial decision-making at the individual level. Lo and Repin (2001) use a novel approach to investigate the role of feelings in investor decision-making. They attached biofeedback equipment to 10 professional financial derivatives traders. With this equipment they collected information on the physiological characteristics associated with emotional reactions, such as sweating and heart palpitations. Lo and Repin (2001) find that traders have heightened emotional arousal around economically important events, such as increased price volatility. The authors do not investigate how this relates to trading performance. Seo and Barrett (2007) do look at this. In an experimental study they let 101 investors participate in a stock investment simulation. For 20 consecutive days the investors had to rate their feelings on a website while making investment decisions. The authors find that investors who experience more intense feelings achieve higher decision-making performance. Furthermore, individuals who were better able to identify and distinguish among their current feelings achieved higher decision-making performance through their enhanced ability to control the possible biases induced by those feelings (Seo and Barrett, 2007). An experimental study from Au et al. (2003) looks at the specific influence of a good or bad mood on trading performance. The authors manipulated the mood state of the participants (all business students) and looked at how this influenced the participants’ trading performance on a simulated foreign exchange trading platform. The authors find that traders in a good mood had an inferior trading performance compared to those in a bad mood. Subjects in a good mood made less accurate decisions and were overconfident in taking unwarranted risks (Au et al., 2003).
9 Grable and Roszkowski (2008) also address the influence of mood on financial risk tolerance at the individual level. Using survey data, they find that people who are in a happy mood display a higher level of financial risk tolerance when confronted with hypothetical investment decisions. The subject of their study is similar. However, the methodology of this thesis is significantly different. Grable and Roszkowski (2008) have a relatively small cross-sectional sample (N=460). This thesis uses a much larger panel sample. Also, this thesis measures happiness and risk tolerance with different questions. Guven (2012) uses the same dataset as this thesis. He investigates the relationship between consumption behavior and happiness. He uses sunshine as an instrument for happiness. Guven (2012) finds that happy people save more, spend less and have a lower marginal propensity to consume. The author does not explicitly look at risk tolerance, but his results clearly show that happy people display different financial behavior. This thesis differs from Guven (2012) by explicitly addressing risk tolerance. Furthermore, happiness is not measured through sunshine, but with a self-reported happiness measure.
2.2 Hypothesis development
All and all, the relationship between mood and financial risk tolerance is not well established and more research is needed (Ackert et al., 2003). This thesis addresses the gap in existing literature by investigating the influence of happiness on individual’s financial risk tolerance. It aims to answer the following question: do happy people have a different financial risk tolerance? Existing literature provides clear support for the idea that happiness, as a mood, influences decision-making under conditions of risk and uncertainty. However, it is not clear whether happiness has a positive or negative effect on risk tolerance. Based on the theories put forward by the psychology literature, two alternative hypotheses are formulated:
Hypothesis 1a: happiness increases financial risk tolerance
Following the line of reasoning provided by the affect infusion model (Forgas, 1995), happiness, as a positive affective state, would increase risk tolerance. Happy moods cue positive memories, which lead to a more favorable assessment of the environment. Some of the existing finance literature on mood and risk taking supports this view (e.g. Grable and Roszkowski, 2008; Au et al., 2003; Hirshleifer and Shumway, 2003).
Hypothesis 1b: happiness reduces financial risk tolerance
10
3 Data and methodology
3.1 Data
This study uses the DNB Household Survey (DHS) dataset, which was formerly known as the CentER Savings Survey. It is a Dutch panel survey among 2,000 households that started in 1993. Each year, the survey collects information on household finances and on economic and psychological aspects of financial behavior. The dataset lends itself well to investigate the influence of happiness on financial risk tolerance, since it combines economic and psychological concepts. Data is currently available for the period 1993-2013. The DHS consists out of six distinctive questionnaires:
1. General household information 2. Work and pensions
3. Housing and mortgages 4. Income and health 5. Assets and debts
6. Economic and psychological concepts
The DHS dataset is very comprehensive and has a complex structure. Response is stored separately per questionnaire, per year. Additionally, there are two aggregated datasets available for each year. The first one includes aggregated data on income. The second one includes aggregated data on assets, liabilities and mortgages. The dataset currently consists out of 160 distinct data files.
This thesis uses questions from four different questionnaires1, for all available years. In other words: it uses data from 80 different data files. All these files had to be combined manually into a new dataset. Every year, the DHS alters some of the questions. Therefore, some of the variables had to be recoded before they could be combined (e.g. education). In 1993 and 1994 one of the control variables (financial knowledge) was not available. These years are excluded. All other cases with missing data on one of the needed variables are deleted as well. For some variables (age and gender) it was possible to retrieve data, based on answers in other years. The resulting dataset contains all relevant variables from the different questionnaires, for all available years.
The final sample consists out of 20,376 person-year observations and covers the period 1995-2013. Within this sample, there are 6,724 unique respondents. On average, respondents stay in the panel for three years. People who leave the panel are replaced. The final sample has an unbalanced panel structure, due to missing observations and panel attrition. Teppa and Vis (2012) show that attrition of DHS panel members is about 20-25% per year. Table 1 describes the sample selection procedure.
Table 1:Sample selection procedure
Initial sample 52,332 person-year observations
Minus: Missing data on happiness 6,607 person-year observations Minus: Missing data on risk tolerance or control variables 25,349 person-year observations
Final sample 20,376 person-year observations
1
11
3.2 Methodology
The properties of the dataset have important implications for the appropriate method of analysis. First, the dataset has a panel structure. The dataset includes repeated observations of the same respondents. OLS assumes that these observations are independent from each other. As this is not the case, it is likely that OLS leads to biased results (e.g. too small standard errors). Panel data techniques, such as fixed effects or random effects, are more suitable (Brooks, 2008).
Second, the dependent variable (risk tolerance) has an ordinal structure. The risk tolerance measure is derived from questions with ordinal answering scales. Again, OLS likely leads to biased results (e.g. the data is often not linear). Limited dependent variable (LDV) models, such as probit or logit, are more appropriate (Brooks, 2008). For ordered response, as is the case in this thesis, generalizations of logit and probit models are used. These are known as ordered logit and ordered probit. The choice between probit and logit is usually arbitrary. This thesis uses an ordered logit model. Traditionally, the logit model was preferred due to ease of computation (Brooks, 2008).
LDV models are usually estimated with cross-sectional data. Panel data techniques assume that the dependent variable is continuous. This dataset requires a combination of both techniques. Combining them is statistically complex, and available estimation methods are limited. It is not possible to estimate an ordered logit model with fixed effects within the available statistical software packages (SPSS, Eviews and Stata). Recently ordered logit with random effects has become available in Stata (version 13). Therefore, this thesis uses an ordered logit model with random effects.
Random effects are appropriate when entities in the sample can be thought of as having been randomly selected from the population (Brooks, 2008). In line with this, the DHS panel is designed to offer an accurate reflection of the Dutch-speaking population (Teppa and Vis, 2004). Contrary to fixed effects, random effect models do not remove explanatory variables that do not vary over time. This is useful for this research, as some of the control variables vary not or only little over time (e.g. gender). Additionally, they estimate fewer parameters, which saves degrees of freedom, and leads to a more efficient estimation than the fixed effects approach. However, the random effects approach is only valid when the composite error term is uncorrelated with all of the explanatory variables (Brooks, 2008). This assumption is more stringent than for the fixed effects case. While acknowledging this drawback, this thesis uses random effects. It simply is the only way to take panel effects into account. This thesis uses the ordered logit model with random effects to estimate the following equation:
12
3.3 Variable measurement
The dependent variable is risk tolerance. The questionnaire contains six statements about investment strategies. Respondents have to answer on a seven-point scale to what extent they disagree or agree (1 = totally disagree; 7 = totally agree) with the following statements:
1. I think it is more important to have safe investments and guaranteed returns, than to take a risk to have a chance to get the highest possible returns
2. I would never consider investments in shares because I find this too risky
3. If I think an investment will be profitable, I am prepared to borrow money to make this investment
4. I want to be certain that my investments are safe
5. I get more and more convinced that I should take greater financial risks to improve my financial position
6. I am prepared to take the risk to lose money, when there is also a chance to gain money This thesis creates two risk measures from these statements. Kapteyn and Teppa (2011) perform a factor analysis on the six statements and show that there are two underlying constructs. Construct 1 combines question 1, 2 and 4. Construct 2 combines question 3, 5 and 6. These two constructs seem quite logical. The questions behind construct 1 are framed different than the questions behind construct 2. The first set of questions focuses on the negative aspects of risk, whereas the second set of questions focuses on the possible gains. Research shows that people respond to questions differently, depending on how they are framed (Nofsinger, 2013).
To validate the constructs created by Kapteyn and Teppa (2011), this thesis performs a principal components analysis (PCA), with varimax rotation and extraction method based on eigenvalues greater than 1. This method assumes the variables are continuous. However, it is common practice to treat the graded responses as if they are continuous (Hofstee, Ten Berge and Hendriks, 1998). Table 2 displays the results.
Table 2: Principal components analysis
Component 1 (RT) Component 2 (RA)
Risk1 (R) .053 .836 Risk2 (R) .434 .547 Risk3 .718 -.007 Risk4 (R) -.003 .848 Risk5 .789 .014 Risk6 .794 .279 KMO .686
Bartlett’s test (sign. level) .000
Note: (R) denotes statements whose score is reversed for the PCA to make all answers consistent (higher score = higher risk tolerance). Largest factor loadings per component are marked bold.
13 the sum of scores on question 1, 2 and 4. Both measures have values ranging from 3 to 21. For RT a higher score means higher risk tolerance. For RA a higher score indicates higher risk aversion.
The independent variable of this thesis is happiness. The questionnaire contains the following question to measure happiness: ‘All in all, to what extent do you consider yourself a happy person?’ There are five possible answers: 1) very happy, 2) happy, 3) neutral (neither happy nor unhappy), 4) unhappy, and 5) very unhappy. This thesis measures happiness with four dummy variables, each one corresponding with one of the answer options. Very unhappy and unhappy are combined, since very few respondents gave this answer.
This thesis includes the following control variables:
Gender. Research consistently shows that women have a lower preference for risk than men (e.g. Sung and Hanna, 1996; Felton, Gibson and Sanbonmatsu, 2003). Both biological and social-cultural explanations have been proposed to explain these differences (Felton et al., 2003). This thesis includes the influence of gender on risk tolerance with a dummy variable (0 = male; 1 = female). This variable is readily available in the dataset.
Age. Existing literature shows that risk tolerance decreases with age (e.g. Sung and Hanna, 1996; Riley and Chow, 1992). People become more risk averse as they reach retirement and the need for a fixed income increases (Riley and Chow, 1992). This thesis measures age with five dummy variables. Each dummy represents a distinctive age category: <35, 35-44, 45-54, 55-64, and 65+. The variable is derived from the birth year of the respondents.
Marital status. There are studies that document a relationship between marital status and risk tolerance (e.g. Sung and Hanna, 1996). However, the exact nature of the relationship is not clear. On one hand, single people may be more risk tolerant because they have fewer responsibilities. On the other hand, married individuals may be more risk tolerant because of a greater capacity to absorb unfavorable outcomes (Hallahan, Faff and McKenzie, 2003). The dataset contains a question about marital status, with six possible answers: 1) married or registered partnership (including separated), having community of property, 2) married or registered partnership (including separated), with a marriage settlement, 3) divorced from spouse, 4) living together with partner (not married), 5) widowed, and 6) never married. This thesis includes marital status with a dummy variable (0 = not married; 1 = married). The dummy variable has a value of 1 for answer options 1) and 2).
14 to evaluate risk (Hallahan et al., 2003). This thesis measures educational level with two dummy variables. They represent the two highest attainable educational levels in the Netherlands: HBO (vocational colleges) and WO (university). The variables are derived from a question about the highest level of education attained.
Occupation. Self-employed people tend to be more risk-tolerant (Sung and Hanna, 1996). Additionally, people in a professional or managerial occupation tend to be more risk tolerant (Sung and Hanna, 1996). This thesis includes the influence of occupation with three dummy variables. First, there is a dummy variable that measures whether a person is self-employed. Unfortunately, the dataset does not tell whether respondents’ occupation is of a professional or managerial nature. However, the dataset does distinguish between working at the government and working at private limited companies (Ltd. / BV). Therefore, this thesis includes two dummy variables representing these two types of jobs, as alternative measures of occupation.
Financial knowledge. There is a positive relation between financial knowledge and risk tolerance
(Grable, 2000). People with more financial knowledge, have a higher confidence level, which makes it easier to accept risk taking (Wang, 2009). The dataset contains one question measuring financial knowledge: ‘How knowledgeable do you consider yourself with respect to financial matters?’ There are four answer options: 1) not knowledgeable, 2) more or less knowledgeable, 3) knowledgeable, and 4) very knowledgeable. This thesis measures financial knowledge with four dummy variables, each one corresponding with one of the answer options.
Table 3 summarizes the variable measurement. Table 3: Variable measurement
Variable Categories Measurement
Risk tolerance RT Combination of risk statement 3, 5, 6
RA Combination of risk statement 1, 2, 4
Happiness Very happy Dummy variable per category
Happy Neutral (Very) unhappy
Gender Dummy variable (1 = female)
Age <35 Dummy variable per category
35-44 45-54 55-64 65+
Marital status Dummy variable (1 = married)
Income Logarithm of net income
Education Vocational college Dummy variable per category
University
Occupation Government Dummy variable per category
Ltd.
Self-employed
Financial knowledge
No knowledge Dummy variable per category
15
4 Empirical results
4.1 Descriptive statistics
Table 4 displays the descriptive statistics of the dataset. It shows that the two risk measures, RT and RA, differ with respect to their mean values. Both have values ranging from 3 to 21. For RT a higher score indicates higher risk tolerance, while for RA a higher score indicates higher risk aversion. The average RA score is higher. This confirms that the two measures are different constructs, measuring different aspects of risk preferences. Regarding happiness, most of the respondents are either happy or very happy.
Table 4: Descriptive statistics
Mean Median Minimum Maximum Std. dev. N
RT 7.831 8.000 3.000 21.000 3.802 20,376 RA 14.962 15.000 3.000 21.000 3.970 20,376 Happiness Very happy 0.203 0.000 0.000 1.000 0.402 20,376 Happy 0.653 1.000 0.000 1.000 0.476 20,376 Neutral 0.133 0.000 0.000 1.000 0.339 20,376 (Very) unhappy 0.011 0.000 0.000 1.000 0.105 20,376 Gender 0.387 0.000 0.000 1.000 0.487 20,376 Marital status 0.708 1.000 0.000 1.000 0.454 20,376 Age <35 0.141 0.000 0.000 1.000 0.348 20,376 35-44 0.202 0.000 0.000 1.000 0.402 20,376 45-54 0.226 0.000 0.000 1.000 0.418 20,376 55-64 0.210 0.000 0.000 1.000 0.407 20,376 65+ 0.221 0.000 0.000 1.000 0.415 20,376 Education Vocational college 0.292 0.000 0.000 1.000 0.455 20,376 University 0.175 0.000 0.000 1.000 0.380 20,376 Occupation Government 0.208 0.000 0.000 1.000 0.406 20,376 Ltd. 0.360 0.000 0.000 1.000 0.480 20,376 Self-employed 0.053 0.000 0.000 1.000 0.224 20,376 Financial knowledge No knowledge 0.184 0.000 0.000 1.000 0.387 20,376 Some knowledge 0.557 1.000 0.000 1.000 0.497 20,376 Knowledgeable 0.224 0.000 0.000 1.000 0.417 20,376 Very knowledgeable 0.035 0.000 0.000 1.000 0.185 20,376 Income(log) 4.259 4.346 -0.032 6.064 0.409 20,376
16 distribution between happiness categories within a dummy variable. Thus, 19.6% of the men are very happy. The second percentage, displayed between brackets, translates this percentage to the entire sample. Thus, 12.0% of all respondents are male and very happy. The last column of the table shows the sum of these second percentages. This is the percentage of respondents with a value of 1 for that specific dummy variable. Thus, for men: 61.3% of the entire sample is male. This is translated to an absolute number between the brackets. For men: 61.3% of the sample equals 12,483 respondents. Finally, for three variables (RT, RA and income) the mean value per category of happiness is given, instead of percentages. For these variables, the last table column shows the overall mean value.
Table 5: Variable distribution per category of happiness Very
happy Happy Neutral
(Very) unhappy % Total (N) Gender Men 19.6 (12.0) 66.1 (40.5) 13.0 (8.0) 1.3 (0.8) 61.3 (12,483) Women 21.4 (8.3) 64.1 (24.8) 13.7 (5.3) 0.7 (0.3) 38.7 (7,893) Marital status Married 23.5 (16.7) 66.5 (47.1) 9.4 (6.7) 0.6 (0.4) 70.8(14,435) Not married 12.4 (3.6) 62.6 (18.2) 22.7 (6.6) 2.3 (0.7) 29.2(5,941) Age < 35 24.5 (3.5) 63.6 (9.0) 10.5 (1.5) 1.4 (0.2) 14.1 (2,874) 35-44 20.3 (4.1) 65.9 (13.3) 12.4 (2.5) 1.5 (0.3) 20.2 (4,124) 45-54 18.6 (4.2) 65.1 (14.7) 15.1 (3.4) 1.3 (0.3) 22.6 (4,611) 55-64 18.2 (3.8) 65.3 (13.7) 15.7 (3.3) 0.9 (0.2) 21.0 (4,269) 65+ 21.3 (4.7) 66.3 (14.6) 11.7 (2.6) 0.6 (0.1) 22.1 (4,498) Education Vocational college 19.5 (5.7) 67.3 (19.7) 11.9 (3.5) 1.3 (0.4) 29.2 (5,953) University 20.0 (3.5) 65.9 (11.5) 13.1 (2.3) 1.0 (0.2) 17.5 (3,571) Other 20.8 (11.1) 64.1 (34.1) 14.1 (7.5) 1.0 (0.5) 53.3 (10,852) Occupation Government 19.0 (4.0) 66.6 (13.9) 13.4 (2.8) 1.0 (0.2) 20.8 (4,246) Ltd. 20.0 (7.2) 65.7 (23.6) 13.2 (4.7) 1.1 (0.4) 36.0 (7,326) Self-employed 21.6 (1.1) 62.5 (3.3) 14.9 (0.8) 1.0 (0.1) 5.3 (1,075) Other 21.1 (8.0) 64.7 (24.5) 13.0 (4.9) 1.2 (0.5) 37.9 (7,729) Financial knowledge No knowledge 20.1 (3.7) 61.1 (11.2) 17.3 (3.2) 1.5 (0.3) 18.4(3,743) Some knowledge 18.4 (10.2) 66.5 (37.1) 14.0 (7.8) 1.1 (0.6) 55.7 (11,350) Knowledgeable 23.4 (5.2) 66.6 (14.9) 9.2 (2.1) 0.9 (0.2) 22.4 (4,561) Very knowledgeable 31.6 (1.1) 60.5 (2.1) 7.5 (0.3) 0.4 (0.0) 3.5 (722) Income(log)* 4.25 4.26 4.25 4.26 4.26 RT* 7.39 7.95 7.89 8.36 7.83 RA* 15.15 14.90 15.06 14.15 14.96
Note: This table shows the distribution of all dummy variables, per category of happiness, with two
percentages. The first percentage looks at the distribution within a dummy variable. The second percentage translates this to the whole sample. For variables marked with * mean values are displayed.
17 clearly shows that married people are happier. Married respondents state more frequently that they are (very) happy. Age is well distributed within the sample. Each age category represents about 20% of the sample. Respondents in the first category, younger than 35, most frequently document that they are very happy. After that age category, the percentage of respondents that is very happy gradually decreases per age category. The percentage increases again, after the age of 65. Within the sample, more people attended a vocational college, compared to university education. People who attended the university answer slightly more often that they are very happy. However, people with ‘other’ education most frequently answer that they are very happy. Compared to the other educational levels, people who attended a vocational college most often answer that they are happy. Looking at the three job categories, it is most common to be employed at a private limited company. It is least common to be self-employed. People who are self-employed most frequently answer that they are very happy. However, they also answer less frequently that they are happy, which translates into a higher percentage for neutral or unhappy. After the self-employed people, people from the occupation category ‘other’ (e.g. retired or unemployed) are most happy. Most people answer that they have some financial knowledge. Relatively few people answer that they are very knowledgeable with regards to financial matters. However, these very knowledgeable people most frequently answer that they are very happy. There are no remarkable income differences between the categories of happiness.
Finally, both risk measures indicate that risk tolerance is lowest for very happy people. The mean value of RT is the lowest for very happy people. The lower this value, the lower the risk tolerance. The mean value for RT is the highest for (very) unhappy people. This indicates that risk tolerance increases, as happiness decreases. On the contrary, the mean value of RA is the highest for very happy people. The higher this value, the higher the risk aversion (or: lower the risk tolerance). Also, the mean value for RA is the lowest for (very) unhappy people. This indicates that risk aversion decreases, as happiness decreases.
Table 6 displays, on the next page, the correlation matrix. It displays the Pearson correlation coefficient for each combination of variables. After each coefficient, it shows the corresponding significance level. All correlations between the independent variables and control variables higher than 0.5 are marked bold. The correlation matrix provides some initial insight into the proposed relationships between the variables.
18
Table 6: Correlation matrix
RT RA
Very
happy Happy Neutral
Very unhappy Gender Age: <35 35-44 45-54 55-64 65+ Marital status Voc. college University Govern- ment Ltd. Self-empl. Income (log) No knowl. Some knowl. Knowled-geable RT 1.000 RA -.321** 1.000 Very happy -.058** .024** 1.000 Happy .042** -.023** -.693** 1.000 Neutral .006 .010 -.197** -.537** 1.000 (Very) unhappy .015* -.022** -.053** -.145** -.041** 1.000 Gender -.209** .152** .022** -.020** .011 -.027** 1.000 Age: <35 .084** -.051** .042** -.015* -.033** .011 .115** 1.000 35-44 .089** -.061** .000 .005 -.013 .018** .024** -.204** 1.000 45-54 .054** -.038** -.023** -.003 .028** .009 -.020** -.219** -.272** 1.000 55-64 -.062** .063** -.027** -.001 .037** -.012 -.021** -.209** -.259** -.278** 1.000 65+ -.151** .078** .014 .011 -.024** -.025** -.079** -.216** -.268** -.288** -.274** 1.000 Marital status .014* .032** .125** .037** -.178** -.074** -.099** -.230** -.015* .060** .085** .063** 1.000 Voc. college .022** .003 -.013 .027** -.027** .015* -.027** .034** .022** -.004 -.018** -.028** -.032** 1.000 University .101** -.088** -.003 .005 -.002 -.007 -.097** .072** .032** -.016* -.043** -.033** -.118** -.296** 1.000 Government -.046** .025** -.017* .013 .002 -.004 -.013 -.065** -.031** .009 .021** .055** -.010 .084** .092** 1.000 Ltd. .068** -.041** -.005 .006 -.002 -.003 -.150** .077** .046** -.003 -.032** -.075** .022** -.075** -.096** -.384** 1.000 Self-employed .041** -.008 .008 -.014* .011 -.002 .018* -.038** -.020** -.011 .007 .054** .016* -.019** .042** -.121** -.177** 1.000 Income(log) .111** -.085** -.012 .019** -.012 .000 -.447** -.083** -.005 .047** .007 .021** -.065** .137** .216** .090** .046** -.060** 1.000 No knowledge -.089** .081** -.003 -.042** .056** .019** .144** -.025** .033** .004 -.004 -.010 .001 -.033** -.063** .025** -.033** -.013 -.149** 1.000 Some knowl. -.023** .054** -.053** .028** .023** .001 .026** -.018* -.033** -.012 .018** .040** -.016* -.004 -.052** -.007 .025** .001 -.017* -.532** 1.000 Knowledgeable .074** -.094** .042** .014* -.065** -.013 -.128** .028** .000 .006 -.010 -.020** .018** .029** .082** -.001 -.002 .007 .118** -.255** -.602** 1.000 Very knowl. .084** -.103** .054** -.019** -.033** -.013 -.083** .039** .018** .009 -.019** -.041** -.001 .015* .087** -.032** .006 .012 .091** -.091** -.215** -.103**
19
4.2 Regression results
Table 7 shows the results for the pooled ordered logit model. The model is estimated twice, since there are two dependent variables: RT and RA. Apart from the dependent variable, the sample is identical. This makes it possible to compare results. The pooled logit model does not take into account panel effects. The model does use heteroscedasticity robust standard errors. For all the dummy variables, one category is excluded, to avoid the dummy trap.
Table 7: Pooled ordered logit
RT RA Coefficient Robust std. error Coefficient Robust std. error Happiness Happy 0.341*** 0.033 -0.167*** 0.033 Neutral 0.385*** 0.047 -0.109** 0.047 (Very) unhappy 0.400*** 0.143 -0.350*** 0.120 Gender -0.747*** 0.029 0.536*** 0.030 Marital status 0.189*** 0.029 0.065** 0.029 Age <35 1.115*** 0.044 -0.560*** 0.043 35-44 0.940*** 0.039 -0.533*** 0.038 45-54 0.764*** 0.038 -0.425*** 0.038 55-64 0.349*** 0.039 -0.076* 0.040 Education Vocational college 0.157*** 0.029 -0.033 0.029 University 0.377*** 0.036 -0.277*** 0.035 Occupation Government -0.108*** 0.033 0.071** 0.034 Ltd. 0.096*** 0.030 -0.038 0.029 Self-employed 0.424*** 0.058 -0.141*** 0.054 Financial knowledge No knowledge -0.630*** 0.082 1.006*** 0.074 Some knowledge -0.441*** 0.077 0.823*** 0.069 Knowledgeable -0.308*** 0.079 0.502*** 0.071 Income(log) 0.017 0.037 -0.007 0.038 Observations 20,376 20,376 Pseudo-R2 0.023 0.013 Log-likelihood -50,430.67 -53,600.70
Note: Significance levels: * = 10%; ** = 5%; *** = 1%. For each dummy variable a reference category is excluded: happiness (very happy); gender (male); marital status (not married); age (65+); education (other); occupation (other); financial knowledge (very knowledgeable).
20 exception of the income measure, all control variables are also statistically significant at 1%. The income measure is not significant. The results show that women are less risk tolerant than men. Also, married people are more risk tolerant. Using the oldest age category as reference category, the results show that risk tolerance decreases with age. A higher education increases risk tolerance. People who attended the university are most risk tolerant. Working at the government decreases risk tolerance, whereas working at a private limited company or being self employed, increases risk tolerance. Self-employed people are most risk tolerant. Finally, risk tolerance increases with financial knowledge. People with the highest financial knowledge are most risk tolerant. Overall, the signs of the coefficients are as predicted.
Contrary to RT, RA measures risk aversion. A higher score on this variable indicates lower risk tolerance. Therefore, coefficients that are positive with RT should be negative with RA. The results with the RA measure show that very happy people have the highest risk aversion. Risk aversion is the lowest for (very) unhappy people. All happiness measures are significant at the 1% significance level, except for ‘neutral’. This measure is significant at 5%. Again, these results are in line with hypothesis 1b. With regards to the predicted signs of coefficients, the results for the control variables are similar to the RT model. With one exception: marital status. According to the second model, married people have higher risk aversion. There are also some differences with regards to significance. Three control variables in the RA model are not significant: working at a private limited company, income and vocational college. Furthermore, some variables are not significant at 1%, but at 5%: marital status and working for the government. Additionally, the age category 55-64 is only marginally significant at 10%.
Overall, the main results are similar. Both measures of risk tolerance show results that are in line with hypothesis 1b. Very happy people have the lowest risk tolerance. The RT measure seems to be a better fit with the data. It may capture risk tolerance better, as more of the results are significant. In line with this, the pseudo-R2 and the log-likelihood are higher for the RT model. The pooled ordered logit model does not take into account panel effects. Including panel effects would be appropriate, as the sample has a panel structure. Due to the nature of the dependent variable, it is only possible to include random effects. Table 8 displays, on the next page, the results for the ordered logit model with random effects. Again, the model is estimated twice, for both measures of risk tolerance. It was not possible to calculate the pseudo-R2 for these models.
21
Table 8: Ordered logit with random effects
RT RA
Coefficient Std. error Coefficient Std. error Happiness Happy 0.297*** 0.046 -0.190*** 0.045 Neutral 0.315*** 0.070 -0.167** 0.068 (Very) unhappy 0.486*** 0.170 -0.279* 0.166 Gender -1.267*** 0.065 0.820*** 0.065 Marital status 0.163*** 0.057 0.043 0.057 Age <35 1.774*** 0.083 -0.950*** 0.083 35-44 1.554*** 0.074 -0.847*** 0.073 45-54 1.299*** 0.069 -0.570*** 0.068 55-64 0.641*** 0.060 -0.174*** 0.059 Education Vocational college 0.182*** 0.062 -0.033 0.062 University 0.341*** 0.078 -0.283*** 0.080 Occupation Government -0.085 0.062 0.044 0.062 Ltd. 0.095* 0.049 -0.046 0.048 Self-employed 0.339*** 0.089 -0.219** 0.088 Financial knowledge No knowledge -0.589*** 0.112 0.877*** 0.110 Some knowledge -0.389*** 0.104 0.716*** 0.102 Knowledgeable -0.274*** 0.101 0.482*** 0.099 Income(log) -0.075 0.056 -0.012 0.056 Observations 20,376 20,376 Log-likelihood -47,200.93 -49,997.93 Sign. LR test 0.000 0.000
Note: Significance levels: * = 10%; ** = 5%; *** = 1%. For each dummy variable a reference category is excluded: happiness (very happy); gender (male); marital status (not married); age (65+); education (other); occupation (other); financial knowledge (very knowledgeable).
22 from 1% to 5%. Finally, the influence of marital status and working at the government become insignificant.
23
5 Discussion
This section discusses the results from the previous chapter. Table 9 summarizes the findings of all four models. For each variable it shows the direction of the relationship with risk tolerance (for dummy variables: compared to the reference category) and whether this relationship is statistically significant.
Table 9: Overview results
RT RA Ordered logit Random effect ordered logit Ordered logit Random effect ordered logit Happiness Happy + *** + *** - *** - *** Neutral + *** + *** - ** - ** (Very) unhappy + *** + *** - *** - * Gender - *** - *** + *** + *** Marital status + *** + *** + ** x Age <35 + *** + *** - *** - *** 35-44 + *** + *** - *** - *** 45-54 + *** + *** - *** - *** 55-64 + *** + *** - * - *** Education Vocational college + *** + *** x x University + *** + *** - *** - *** Occupation Government - *** x + ** x Ltd. + *** + * x x Self-employed + *** + *** - *** - ** Financial knowledge No knowledge - *** - *** + *** + *** Some knowledge - *** - *** + *** + *** Knowledgeable - *** - *** + *** + *** Income(log) x x x x
Note: +/- denotes sign of relationship with dependent variable. Significance levels: * = 10%; ** = 5%; *** = 1%. ‘X’ denotes insignificant relationships. For each dummy variable a reference category is excluded: happiness (very happy); gender (male); marital status (not married); age (65+); education (other); occupation (other); financial knowledge (very knowledgeable).
24 model with RA, two happiness variables have a lower significance level: ‘neutral’ (5%) and ‘(very) unhappy’ (10%).
In the RT models, differences in risk tolerance are most pronounced between the very happy people and (very) unhappy people. It is least pronounced between happy and very happy people. Thus, risk tolerance gradually increases as happiness decreases. The RA models provide a somewhat different picture. Again, the difference in risk aversion is the largest between the most happy and least happy respondents. However, in the RA models the difference is the smallest between people in a neutral mood and very happy people. Thus, people that are neither happy nor unhappy have a smaller risk tolerance than happy people. Risk tolerance actually increases between the happiness categories ‘neutral’ and ‘happy’. This does not change the main result, as both categories still have a higher risk tolerance than very happy people. However, it highlights the complexity of the subject under investigation. Different measures of risk, highlighting different aspects of risk, can alter the results. The RT measure is derived from statements about investment strategies that are framed positively. The questions contain words like profitable, improve, and gain. The RA measure is derived from statements that are framed negatively. These questions focus more on the potential losses and contain words like safe, certain, and never. Risk taking can be investigated along several dimensions, including: outcome uncertainty, outcome expectation, outcome potential, personal involvement, and perceived safety and control (Chou et al., 2007). The questions behind RA seem to focus more on perceived safety and control, whereas the questions behind RT may focus more on outcome potential. Overall, RT seems to be a better measure of risk tolerance. There are virtually no differences between the pooled RT model and the random effects RT model. For the RA models, including random effects lowers significance levels of some variables. Apart from that, the RT models provide significant results on more variables than the RA models.
For most of the control variables, results are as expected. Women displayed a significant lower risk tolerance than men. The variable that measures marital status provided mixed results. Being married increased risk tolerance in both RT models, whereas it led to increased risk aversion in the pooled RA model. In the random effects RA model, the variable was not significant. These mixed results reflect the conflicting findings in existing literature (Hallahan et al., 2003). As expected, risk tolerance decreases with age. People younger than 35 are most risk tolerant, people in the oldest age category (65+) are most risk averse. Risk tolerance increases with educational level. Respondents with university education have the highest risk tolerance. The vocational college variable was not significant in the RA models. People who are self-employed display the highest risk tolerance, compared to other forms of employment. Significance of the other two occupational variables differs across models. Existing literature mentions a positive relation between professional occupations and risk tolerance. The dataset did not provide this kind of information. Therefore, this thesis used working at the government and working at a private limited company as proxies. This might explain the lack of consistent significant findings. As expected, risk tolerance increased with level of financial knowledge. Finally, in none of the models the income variable was significant. This may be caused by its measurement: the measure aggregates different forms of income.
25 this thesis fit within this model, as they show the influence of a specific feeling (happiness) on a specific aspect of decision-making (financial risk tolerance).
Existing psychology literature is divided about the direction of the relationship between a good mood and risk tolerance. According to the AIM, a good mood increases risk tolerance (Forgas, 1995). According to the MMH, a good mood lowers risk tolerance (Isen and Patrick, 1983). The findings of this thesis can be explained by the MMH (Isen and Patrick, 1983). According to the MMH, people in a good mood (or: happy people) are less willing to take risks, because they want to maintain their current positive affective state. Taking risks increases the potential for substantial losses, which are detrimental to a good mood (Chou et al., 2007). On the contrary, people in a bad mood (or: unhappy people) do take high risks, as the potential gains may uplift their mood (Chou et al., 2007).
This thesis differs from psychology literature, by focusing on financial risk tolerance. Risk tolerance is an essential element of financial decision-making. This thesis shows a significant influence of happiness on risk tolerance. These results are in line with the growing number of finance articles that acknowledge the influence of mood on financial decision-making (e.g. Kamstra et al., 2003; Hirshleifer and Shumway, 2003; Kliger and Livy, 2003; Lo and Repin, 2001; Au et al., 2003). The articles of Guven (2012) and Grable and Roszkowski (2008) most closely resemble the subject of this thesis. Guven (2012) uses the DHS dataset to investigate the relationship between happiness and consumption behavior. The author finds that that happy people behave differently. This thesis extends these findings by focusing on risk tolerance.
26
6 Conclusions
This thesis started by comparing traditional finance literature with Star Trek. Finance literature seems to assume our world is populated by Vulcans, a species that suppresses emotions. Inspired by behavioral finance literature, this thesis diverged from the Vulcan-assumption and investigated the influence of happiness on financial risk tolerance. So far, most research on financial risk tolerance ignored the influence of feelings (notable exception: Grable and Roszkowski, 2008). Using data from the DNB Household survey, this thesis finds a negative relationship between happiness and financial risk tolerance. People who report the highest level of happiness, display the lowest level of risk tolerance. These findings are consistent for two different measures of risk tolerance, across different model specifications.
Overall, the findings contribute to a broader understanding of the influence of feelings in financial decision-making. They add to the growing finance literature that acknowledges the influence of feelings. Contrary to the existing finance literature on mood and decision-making, this thesis uses micro-economic data and focuses on households. This thesis distinguishes itself from existing psychology literature, by addressing risk tolerance in a financial context, and by using large-scale survey data instead of an experimental design. A better understanding of the influence of feelings has practical relevance for personal financial planning and investment management. Assessing capacity for and attitude towards risk is the key to successfully implementing an investment policy (Filbeck, Hatfield and Horvath, 2005). This thesis shows that feelings are a factor that should be taken into account in the assessment of risk tolerance.
Like any other research, this thesis faced several limitations. This thesis uses a pre-existing dataset. This determines what variables are included and how they are measured. The DHS dataset is not designed with the specific purpose to investigate the relationship between happiness and financial risk tolerance. With a questionnaire specifically designed for this subject, some variables would have been measured differently. For example, it would contain more detailed questions about respondents’ occupation. Additionally, the dataset did not include information on other moods. Investigating the influence of multiple moods at once provides a richer picture of the influence of feelings on financial decision-making. However, this limitation does not outweigh the benefit of using the very extensive DHS dataset. The dataset allowed to investigate the subject with a large sample (20,376 person/year observations) covering a long period (1995-2013). Another limitation follows from the sample itself. The sample has a panel structure and has an ordinal dependent variable. This combination limits available methods of analysis. It is not possible to estimate an ordered logit model with fixed effects. This thesis uses random effects as an alternative. As a result, panel effects may not have been accounted for in the most efficient way. Therefore, some caution is warranted in interpreting the results.
