Relative income versus absolute income : a case study : can the one, help replace the other?

(1)

1 Relative Income versus Absolute Income: A Case Study. Can the one, help replace

the other? Ryan Bond 10141758

Behavioral Economics and Game Theory 12 ECTS

Supervisor: Frans van Winden University of Amsterdam

Abstract

One promising solution for the Easterlin Paradox, which says that more income does not mean more happiness, is using relative income instead of absolute income. This paper attempts to say something about the use of both by looking at correlations and differences therein. Previous research found links between trait-self-control, absolute income and happiness, here we check these links for absolute and relative income.

Relative income is measured with a subjective measure, as in Mangyo and Park (2011), which has not previously been done in happiness research. The paper also does something new by testing happiness measures with different perspectives to see

if there are differences in quality and use. Instrumental variables are used where possible to alleviate endogeneity problems such as reverse causality and omitted variable bias. The paper uses its own dataset, leaving out the data gathered from India, because of possible measurement error problems. Results indicate that relative

income predicts better or at least as good compared to absolute income in various models and that a happiness measure that compares happiness with others instead of

being just an absolute happiness measure shows strong correlations with the regressors. This paper further considers the issue of causality by doing a mediation

(2)

2 Statement of Originality

This document is written by Student Ryan Bond who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in

creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

1. Introduction

Most common in economics is the concept of the homo economicus, which assumes that agents maximize their utility in every economic decision. Being rational beings, we maximize payoffs usually denoted in monetary units, which seems straightforward and logical. Except that economists have already found countless pieces of evidence against this most important assumption in agents. More is not always better, and humans are not purely selfish. A nice example is the work of Easterlin (2005), where it is found that when all incomes double, nobody gets any happier. What then is the use of this assumption in economics? Policy makers decide what’s best for the public by maximizing the ‘common good’. Most economists agree with this argument, but knowing these

contradictions, it may be time to change the assumptions at the core of economics. The ‘common good’ may not be so common after all.

Happiness Economics is a relatively new body of literature, it tells us that agents do not focus only on income to calculate utility, as is assumed in regular economics. Instead, they use happiness, well-being or life satisfaction, as a substitute for utility (Frey, 2008). Happiness, which this paper will focus on, is

(3)

3 contrived by many factors that economists should identify. This strand of

literature sprung into existence due to the Easterlin Paradox, which basically says that more income does not mean more happiness (Easterlin, 2001; Stevenson & Wolfers, 2008). This paradox has been investigated thoroughly and has even been partially or fully ‘solved’ in a few papers (Di Tella & MacCulloch, 2008). One fix uses relative income instead of absolute income as a determinant of happiness (McBride, 2001). This is promising because as Easterlin (1995, p.36) argues, increasing everybody’s income would not make more happiness, because prices would go up as well. Increasing relative income however, would not have this problem, thereby solving the paradox. Papers that corroborate the finding that income of reference groups matter are plentiful (Ferrer-i-Carbonell, 2002; Mangyo & Park, 2011). As can be expected, poorer individual’s happiness is negatively influenced by the fact that their income is lower than that of their reference

group(s) (Ferrer-i-Carbonell, 2002). Ferrer-i-Carbonell also finds that this does not inversely hold for richer individuals, meaning that comparisons are mostly

looking up to richer people (2002).

Income has traditionally been measured in currency alone, or as an absolute measure. Recent developments find evidence for another option. Easterlin (2001), finds that relative income (RI) instead of absolute income (AI) influences

happiness. These two options, AI and RI, have been the two sides of the discussion that is going on.

About as long, there has been a discussion about whether relative income should affect subjective well-being or happiness (Yu & Chen, 2016). This

originates from the belief that people compare situations/options to value them, to subsequently choose between these options (Fitzgerald, Friston & Dolan, 2012). The relative income hypothesis hinges on the assumption that people compare to other people. Social comparison is regarded as a mechanism for joining income and happiness and without social comparison, relative income shouldn’t affect happiness (Deaton and Stone, 2013).

(4)

4 Yu and Chen (2016), find that RI and AI weaken negative emotions but only RI improves positive emotion. This could imply that absolute income has nothing to do with happiness. Absolute income is implicated in several measures of well-being in the literature though. However, some of these measures do not only measure happiness, but also satisfaction, which might be different things or the same, depending on how the questions are framed and how people think about them. This suggest that RI says more about happiness than AI. Furthermore, Deaton and Stone (2013), found that relative income was significantly associated with happiness, but also that this was not so for absolute income. This is directly in line with what Easterlin (2001) and Mcbride (2001) found.

Alpizar, Carlsson and Johansson-Stenman (2005), argue that people care about both absolute income and relative income. For the income measures, what they found is that on average, 45% of the utility increase from extra income comes from enjoying a higher relative income (Alpizar et al, 2005). This shows that not just one of the two income measures but rather both influence happiness and gives a perspective for looking at how the measures influence happiness. The issue of how important relative income is for happiness compared to absolute income is hard to answer, but it shows that it’s not just one of the extremes but rather a blend of the two.

Agents try to maximize utility and this paper uses happiness as a measure of utility. Happiness is usually measured with a focus on how satisfied one is with life, however, there is another way that measures happiness compared to others. Asking subjects to compare their happiness to others on a scale of 1, less happy, to 7 more happy. The usual happiness measures are about absolute happiness and this alternate one is more about relative happiness. The difference between the two happiness measures mirror the situation with AI and RI. This paper will also compare the two to see if the way of measuring happiness influences the results.

(5)

5 The paper uses its own dataset comprised of subjective data on age, sex,

working hours, addictions, RI, AI, happiness and trait-self-control gathered using Qualtrics questionnaires, one in English and one in Dutch. Distributed via Mturk and Facebook respectively. This will be discussed later. The choice of using a new dataset was one of necessity and expediency, since the plans required various data that don’t appear in any known dataset, particularly data on TSC and RI. What’s more is that using a questionnaire was a cost- and time-efficient way of gathering the required data cheaply and quickly. This supplied data from 354 usable subjects, of which only 243 were used in the main section. This was

because of problems with the data from India. Of the 243, 120 of these were from the USA and 123 from the Netherlands.

Along with checking the differences of the correlations and causations

between RI and AI with happiness, this paper also uses another factor with ties to happiness and income. Trait-self-control, or TSC, is a determinant of both

happiness and income (Cheung, Gillebaart, Kroese & De Ridder, 2014; Hofmann, Luhmann, Fisher, Vohs, & Baumeister, 2014). The causality between TSC,

happiness and income isn’t fully clear though, so that must be considered (Stutzer & Frey, 2006). That will be done using measures such as IV-estimation to offset reverse causality issues and a mediation analysis. This paper will use this TSC factor to test the different income measures further, as it relates to happiness, or utility.

Absolute and relative income influence happiness along with TSC (Moffit et al. 2011; Diener & Kesebir, 2013; Cheung et al, 2014). This paper investigates these three relationships by looking at correlations, causations and differences therein. Differences between the two happiness measures that are used in the paper lead to some additional insights that might influence the results. Using subjective measures of RI has not been done in the literature before and maybe it will tell economists a little about if and how happiness economics can be integrated into

(6)

6 regular economics. If for example the correlations between RI and TSC together with the correlations between RI and Happiness are greater than those of AI with TSC and Happiness, that would prove that AI is not all that agents are trying to optimize and that RI is also to be considered. This would not render the economic assumption of maximizing obsolete, just the utility measure of absolute income that is generally used. Therefore, the results will tell us something about the possibility for RI to replace or to help replace AI as an income measure related to utility. This will reinforce what the Easterlin paradox told us, namely that the AI measure is not the complete answer. Leading to an answer to the following question: Can relative income help replace absolute income? And if so, to what degree?

In the next section, the dataset will be described. Then in section 3, the empirical methodology is described. Section 4 will be the result section, divided into 3 subsections. 4.1 will be about the models regarding happiness and will detail the differences found between RI and AI and between the two happiness measures. 4.2 will do the same for the models regarding TSC and 4.3 will show the results of a mediation analysis a la Baron & Kenny (1986). Lastly, section 5 will host the discussion and conclusion.

2. Data

To gather enough data in a timely fashion I used an anonymous questionnaire that was made with Qualtrics. This questionnaire was distributed via Facebook and Amazon Mturk. The latter is an on-demand workforce, meaning that for small amounts of money, workers from around the world do tasks for requesters, such as businesses and students that need data. Using Mturk for gathering data is a proven method (Fowler, Willis, Moser, Ferrer & Berrigan, 2015). Fowler et al. (2015) were surprised to see how good the results were, given the low extrinsic motivation or payments involved.

(7)

7 The number of subjects totaled 423, of which 248 were gathered using Mturk and 175 using Facebook. 69 had to be dropped, meaning only 354 were usable, of which 123 were gathered using Facebook and 231 using Mturk. Dropping subjects was done for assorted reasons such as cheating (on Mturk), and being outside the age limit (20-65 years old). For a full accounting of this vetting process, see

Appendix A, which holds the vetting process together with the models used in this paper.

Questionnaire

The questionnaire was designed in Qualtrics using existing scales and questions for gathering subjective measures on Happiness, TSC and RI. Every question was answered by all subjects since they could not finish the

questionnaire otherwise. This was possible because of a feature of Qualtrics and ensured no missing values.

For measuring Happiness, the General Happiness Scale (4 questions) was used together with the Satisfaction with Life Scale (5 questions). Lyubomirsky &

Lepper (1999) tested the first, General Happiness Scale, and found excellent

reliability and consistency. The Second, the Satisfaction with Life Scale, was tested by Diener, Emmons & Larsen (1985) who found the same, i.e. high reliability and consistency. For measuring TSC, the Brief Self-Control Scale (BSCS), consisting of 13 questions, was used. Lindner, Nagy and Retelsdorf (2015) evaluated the BSCS and recommended its use. The RI measure is based on 7 subjective questions used in Mangyo & Park (2011). Mangyo and Park concluded that using a subjective RI measure gave meaningful results. How the scales work can be found in the literature described above and is also shortly mentioned in appendix B which holds the questionnaire.

Using existing measures with proven usefulness should ensure meaningful results. Besides these measures, the questionnaire asked the subjects about whether each subject is male/female, smokes or not, drinks or does not, does

(8)

8 drugs or not, thinks he/she has an addiction and what they earn (gross yearly income). The addiction-related questions asked the subjects to rate themselves from 1-definitely not, to 5-definitely yes. These points added up to an Addiction measure. All the subjective questions in the entire questionnaire used 5-point Likert scales except for the Happiness measures, which used a 7-point Likert scale. See Appendix B for the full questionnaire.

The survey was divided in two blocks, block 1 reported age, working hours, gender, AI, RI and an addiction measure and block 2 reported the happiness and TSC scales. The second block with the questions regarding the happiness and TSC scales were randomized, to avoid answering biases. In hindsight, it would have been better if every question had been randomized. Since now there still is a possible answering bias caused by the sequence of reporting income before happiness. And another one caused by the sequence of reporting addictions before happiness. This could affect how people have answered the happiness questions and is therefore a weak point of this research.

For the Mturk questionnaire I added a control-question that stated: please select option Very much. There was also an invisible question that timed the subjects while they answered the questions. The latter two questions were solely for making sure the subjects were answering the questions carefully instead of just filling out the form quickly. These were used to decide whether to drop the subjects and not pay them, since requesters don’t have to pay workers on Mturk if they do not perform satisfactorily on their assigned task. The control-question was not included in the Facebook questionnaire since there the incentives were purely intrinsic, they received no compensation, and it was unlikely to find

subject who cheat their way through the questionnaire. Unlike the Mturk workers who mostly do it for the money involved.

The survey collected data anonymously, where the only identifiable

(9)

9 workers and to get the Facebook group to be sufficiently confident to answer truthfully how much their gross yearly income is. The absolute income question asked the subjects to write down their gross yearly income rounded to thousands combined with the currency used. This however was not always understood and many subjects forgot the currency. If not for the IP address that would have made part of the data useless. Though the IP address is quite anonymous in the hands of a mere economist, it has one feature that can be used to figure out roughly where the signal came from in the world. Using this, it was possible to acquire the country information and fill in the missing currency information. This worked well.

Most of the subjects, 330 (93%), were from three countries, India, the USA and the Netherlands. The subjects from the Netherlands were all gathered using Facebook and the rest using Mturk. Which makes sense since Dutch people cannot be workers on Mturk and the majority of Mturk workers consists of people from the US and India.

Comment on the three main countries

The three top countries making up 93% of the data are unequally represented. Checking the descriptive statistics, India deviates strongly from the mean with one of the main statistics, III, which is absolute income in international dollars rounded to thousands. How this variable is calculated is explained in the next subsection called variables. For now, it is only needed to understand that using the international dollar means the currencies have been normalized, which should make them equal. As shown in table 1 below India enjoys less than half the income of the rest of the countries.

Table 1: Absolute income for three main countries variable III

IND III USA III NL III N-IND

Mean 14.06 38.59 31.03 34.30

Obs. 87 120 123 267

(10)

10 This is a very strange result, in fact there is no clear reason why India III is so low compared to III in other countries. On another note, these results could be

because of how the gross income reporting question was framed. The question asked the subjects to report their yearly gross income rounded to thousands, and to report the currency in which their income was denoted if it was not denoted in dollars. This left room for measurement errors. People forgot to mention their denoted currency because they misunderstood the question or they just forgot to. This measurement error was bypassed by using the IP-addresses to find the country of origin. There might still be some who forgot to note the currency after exchanging it for dollars, which were then thought to be in INR due to the IP-address. Though this mistake is doubtful. Either way, there was no way to check if they did because of the high variance in Indian incomes. If for example

someone worked for 40 hours a week and earned 1000 international dollars in the USA, it would be clear to see something is wrong. Some Indians however, earn no more than .30$ an hour, while others earn fifty times that, making it impossible to judge the reliability of the incomes. Therefore, there could be a measurement error there.

The difference in results from India would mean different correlations and could impede the predictive power of the models. Rather than trying to find whether this is really happening it was decided to forego the use of the data from India. For even if these troubles were found, it is impossible to find the true reason. This demonstrates the need for extensively pre-testing a questionnaire, which could not be done properly in preparation for this paper. Because the subjects from the other countries not already mentioned were few, and likely to suffer the same problems, those were also omitted.

The data left consists of 243 subjects, of which 123 came from the Netherlands and 120 from the USA, which are quite similar in number and in culture, i.e. they are both rich and developed countries. This similarity gives some confidence in

(11)

11 the results that would have been lessened or even lacking if subjects from India, and the other countries, which are not rich and developed or notably less so, were to be included.

Having the data come from only two countries also means that 123 subjects, those from the Netherlands, come from the use of Facebook and the other 120 from the USA come from the use of Amazon Mturk. That difference in data also constitutes a difference in incentive. Whereas Mturk workers work for small amounts of money, Facebook friends and friends of those friends filled in the survey from an internal incentive, an intrinsic motivation, without getting recompensed financially. These different perspectives and circumstances could make a difference in how the questions were answered. Specifically, since the Mturk workers were more income-oriented, because by filling in the survey, they were earning money.

However, three reasons may mitigate this effect. Firstly, not randomizing the first block of the questionnaire means that all subjects were first asked about their income and relative income before answering the happiness questions. This would mean the Facebook group was also income-minded when gauging their happiness, as was the Mturk group. Secondly, Fowler et al. (2015) reported that intrinsic motivation was greater than extrinsic motivation for most Mturk workers. This would indicate a smaller difference than expected between the results from Facebook and Mturk. Thirdly, the two groups were about the same in terms of numbers, this should even out the effects of any problems. Because the difference seems to be less significant than first expected, this paper will take the data at face value.

Variables

Because the subjects came from all over the world and earn their incomes in different currencies, using the gross income statistic as a variable was not an option. Differences due to living standards and conversion rates had to be

(12)

12 eliminated. To do this all gross incomes were converted over to international dollars using the implied PPP conversion rate of the World Economic Outlook (WEO) of April 2017 found at the IMF website (World Economic Outlook, 2017). PPP, or purchasing power parity, is a ‘’basket of goods approach’’ to compare incomes across countries and currencies. From then on, all incomes were of one (international) currency and could be used as a variable, labeled variable II. The II variable was divided by a factor of a thousand to make a new variable III for more meaningful results. In table 2 below the descriptive statistics of the main variables are shown. In this table, Female is a dummy equal to 1 for female and 0 for male, Whours is hours worked per week and III is income in international dollars in thousands. RI, Hap1, Hap2, TSC and Addict are the total of their respective scales and age is the age of the subject. Variable Hap1 represents the Satisfaction with Life Scale, which is an absolute happiness variable. Variable Hap2 represents the more relative General Happiness Scale and variable TSC represents the Brief Self-Control Scale.

Table 2 below shows descriptive statistics for the main variables. Table 2: Descriptive statistics for the main variables

Variable Obs. Mean Std. Dev. Min Max

Female 243 .502 .501 0 1 Whours 243 34.634 13.665 0 70 Addict 243 8.144 3.276 4 18 RI 243 20.263 5.125 7 35 Hap1 243 23.230 6.686 5 35 Hap2 243 18.214 5.681 4 28 TSC 243 42.984 9.538 18 65 Age 243 34.420 12.317 20 65 III 243 34.995 22.476 .759 119

Note: refer to text for definition of variables.

All subjects are between 20 and 65 years of age. Slightly more than half of them are females and on average they approach the means of their countries income level per capita. This suggests the sample could well be used to represent reality.

(13)

13 Income and happiness are the crux of this research, and we have two countries represented in the dataset, USA and NL. First, their differences will be shown.

Figure 1 shows how III and RI are spread over the countries.

There’s three things to see here, first, the absolute incomes are higher in the USA. Second, the NL have higher subjective relative incomes and thirdly, the spread is smaller for the NL in both III and RI. This smaller spread could be because the NL is a much smaller country, causing the people to feel more as one community. This trend continues for the Hap1 and Hap2 measures spread over the countries as seen

(14)

14 in Figure 2. Furthermore, the NL seems to enjoy higher happiness than the USA.

These results agree with the current statistics. USA is below the NL in happiness statistics (World Happiness Report, 2017) but above the NL with income statistics (World Economic Outlook, 2017). The people of NL score higher on RI despite scoring lower on III, this could just be because the Dutch are happier in general, since comparing is rooted in happiness and Hap2 is the more relative measure.

The statistics of the whole dataset, comprised of both the USA and the NL, can be found in Appendix C, in tables C1 to C5, which comprise the questions on addictions, RI, TSC Hap2 and Hap1 respectively.

3. Empirical Methodology

Studies testing the relative income effect have mostly the same setup, one that requires panel data. As shown in a comparison paper done by Brown, Gray and Roberts (2015), lots of methods have been used, all of them using panel data, with different time periods. This paper however, uses a single questionnaire that takes just one moment in time, rendering the use of panel-data regression techniques moot. In addition to OLS, this paper uses IV-estimation regressions, which, according to literature has not been done much. This is most likely because of the use of a subjective relative income measure, which has only been done once, by

(15)

15 Mangyo and Park (2011). They used their subjective relative income measure to do OLS regressions. But where Mangyo and Park used their subjective relative income measure to test for health measures, this paper tests for happiness

measures. The difference between the two is that there are very often endogeneity issues found while testing for happiness (Hajek, 2013). The literature provides a lot of evidence linking virtues such as TSC to happiness and income measures, also to each other, meaning these are all interrelated (Kesebir & Diener, 2013).

Causality is the issue here, as it is expected to flow both ways. People know they get happier by having more money, but it could also be the other way around. This has been studied before in multiple papers, drawing differing conclusions. A few of these papers are, Li, Liu, Ye & Zhang (2011), Hajek (2013), De Neve & Oswald (2012), and Oswald, Proto & Sgroi (2009). This paper tries to take the causality into account, and find its direction. Causality however, is difficult to measure with any sort of accuracy. It suffices to say that these

correlations cause issues, such as reverse causality and omitted variable bias, and that OLS-estimation results might not always show significant results in these circumstances. Therefore, other than OLS, this paper will use IV-estimation to counter those endogeneity issues and improve the accuracy of the results.

Making the models

This paper aims to find correlations and causation between AI, RI, TSC and Happiness to be able to make a comparison between AI and RI. Leaving the correlation between AI and RI out that leaves 5 links that must be investigated both ways, making 10 regressions. Regressions 5 and 6 between TSC and Happiness do not concern income but were done regardless because it tells us something about the underlying relations and causation of the variables. In Figure 3 below is depicted what the 10 regressions are and which way the

(16)

16 causation is implied. Arrow 2 for instance, would be a regression in which TSC would explain/predict III with the reverse being true for arrow 1.

Describing all the models would take much space and would be of limited interest here. So instead, in generic terms will be described how the models are made using regression 4 as an example. The full make-up of the models and their instruments will be described in table A1 in Appendix A, right after the vetting process.

To start making models using the variables available its necessary to know if and how the variables correlate and predict each other. So, using regression 4 with Hap2 as the example, III and the rest of the variables were regressed on Hap2. By using stepwise regression backward elimination tactics, i.e. leaving the weakest regressors out and starting over until all the regressors are significant, the insignificant ones were left out, leaving RI, TSC and the main regressor III as regressors for the dependent variable Hap2. So, throughout the paper, always the two variables of interest remained as dependent and explanatory variables. Combining those with additional significant regressors made the models. For regression 4 this makes the following OLS model:

𝐻𝑎𝑝2 = 𝑎₁+ 𝛽₁𝐼𝐼𝐼 + 𝛽₂𝑇𝑆𝐶 + 𝛽₃𝑅𝐼 + 𝜀₁ Instrumental Variables estimation

To use IV-estimation, or 2-stage-least-squares (2SLS), instruments are needed. In our example of regression 4, we need instruments for III because of suspected endogeneity. These instruments need to predict III and not predict Hap2, i.e. be

(17)

17 uncorrelated with the error term. Finding suggestions for these with the correlate function in Stata, a statistical program, was done in the same way as finding additional regressors before, with backward elimination. This process came up with instruments Age and Female. Which makes the following model:

𝐻𝑎𝑝2 = 𝑎₁+ 𝛽₁𝐴𝑔𝑒 + 𝛽₂𝐹𝑒𝑚𝑎𝑙𝑒 + 𝛽₃𝑇𝑆𝐶 + 𝛽₄𝑅𝐼 + 𝜀₁

The process described above was used for all the models and led to valid instruments, meaning the scores from F-tests, or Wald tests, on the validity of instruments where sufficiently high (>10) for all of them. The scores can be found in table 4 and 5 in the results section. Doubt resulting from this process is about the validity of the model. A weak model would not give significant and

meaningful results. The Stata program has ways to test for this though, tests such as the endogeneity test and over-identification test indicated whether the model indeed faced such problems, and if the model was correctly specified.

For testing endogeneity, Stata uses two tests, the Wooldridge test and a regression based test. The regression-based test however, is amenable to

clustering. This dataset might be featuring two such clusters, because there are two distinct groups divided by country and way of data gathering. Typically, for large n, these tests yield similar conclusions. Having a smaller dataset can cause discrepancies between the two tests. In this case, the Wooldridge test will most likely have the larger p-values, making that the more conservative test, i.e. less willing to make conclusions. Therefore, this paper will only report the

Wooldridge test results. But if the Wooldridge test does not reject exogeneity and the regression-based test does, the test is still taken as significant in the results section, because even if the regressor variable is exogenous, the 2SLS estimates should still be consistent. Whereas the OLS estimates are not consistent in the case the regressor variable is endogenous.

Heteroscedasticity, or related error terms, can invalidate statistical tests and coefficients. Using the White-test for heteroscedasticity, every final OLS model

(18)

18 was tested. Of the 20 models, 18 tested positive for heteroscedasticity with less than 10% statistical error, i.e. with a p-value smaller than .10. For all those models with heteroscedasticity, robust standard errors, or the Huber/White/Sandwich estimators were used. These estimators relax assumptions and make the results robust to heteroscedasticity. This does not influence the coefficients found, but only the statistics using variances, such as the standard errors and the statistical tests.

Clustering can be a major issue in some models, though with the models used in this research it is not expected to matter much, this is because clustering

essentially means you get the same information twice, from different subjects. Like asking two friends how many friends they have. However, in this

questionnaire all the questions are subjective and about the self, particularly the questions about traits. Therefore, there is no reason to expect any sort of

clustering except maybe for income, when two subjects work together. Using cluster options in Stata always renders a result, but gives no indication of

statistical validity, which makes it very hard to interpret. Also, there is no rule of thumb for the sample size required or how many cluster variables are needed (Sarstedt & Mooi, 2014). Formann (1984), uses 2𝑚, where m is the number of clustering variables used. With this small dataset, (N=234) this seems a tall order, because with 9 variables this would require 29 = 512 data points. Due to these reasons, it seems better to not use clustering.

Most of the ten IV-regressions use two instruments, this is needed to perform an overidentification test, which tests instruments’ endogeneity together with testing the equation for misspecification. A positive test score means either an instrument that is invalid or it means that one or more of the instruments should be in the equation. Wherever this overidentification test was not possible because there was only one instrument, that instrument was tested for correlation with the error-term, which also measures endogeneity. All these tests gave an p-value of

(19)

19 1.0000, which means there is no connection whatsoever between the instrument and the error term of the model.

To test both happiness measures, all regressions will be done double, one for each measure, making 20 regressions total. As can be seen in table A1 in appendix A, these models are not always the same because the two measures measure happiness differently, they also correlate to the rest of the variables differently. For example, we find that females are happier than men, which can be seen in table 3, but the effect is much bigger for the correlation of Hap1 with female over Hap2. Additionally, the correlation found for Hap2 is nowhere near significant, reinforcing the differences found.

Table 3: Correlations with Corresponding P-values

Female Whours

Hap1 .1627**

(.0111) -.1168* (.0692)

Hap2 .0434

(.5008) -.0007 (.9915) Notes: 1) * significant at 10%; ** significant at 5%.

2) between parentheses are the P-values corresponding with the correlation measures.

This could mean females compare incomes less compared to males. Another example is even more straightforward. Working more hours in a week should decrease happiness, because there is less time for leisure. We see this clearly for Hap1 with a correlation with Whours of -.1168 but not for Hap2, for which the correlation of .0007 is barely different from zero. The corresponding p-value further indicates that there is no relation between Whours and Hap2 while there is for Whours and Hap1.

These examples illustrate quite clearly why substituting Hap1 for Hap2 in models or the other way around is not a sound idea in that it would not fit the same way, leading to less efficient models. Therefore, the models needed to be fit differently for both happiness measures, though they are remarkably similar most of the time regardless as can be seen in table A1 in Appendix A.

(20)

20 The next section will be about the results. This section will be divided in three subsections, the first two being about the models regarding happiness and the models regarding TSC respectively. There the differences between RI and AI, represented by III, and the differences between the two happiness measures will be discussed, together with some remarks about causality. This will be further discussed in the third section, where a small mediation analysis will be done. Section 5 will discuss the results and have some concluding remarks.

4. Results

In this section, for every model, the dependent variable is mentioned, as well the prominent regressor. The additional regressors were not mentioned to save space, though table A1 in appendix A explains all models, including the

additional regressors.

4.1 Differences between III/RI and Hap1/Hap2 for models on happiness.

The most important models in this paper are regarding happiness, if you recall figure 3 that means regressions 3-6 and 9-10. This section will find try to find the answer to the main questions that ask how RI does compared to III and how using a different happiness measure influences results. To answer the first, regressions 3 and 4 will be compared to regressions 9 and 10. To answer the second, all regressions were done double, one with Hap1 and one with Hap2. This makes it possible to compare all regressions done using Hap1 with the ones using Hap2. Where Hap1 was an absolute measure of happiness and Hap2 a more relative measure. Table 4 shows the OLS and IV-estimates for the first 6 models for both happiness measures.

Table 4: OLS and IV-estimates on regressions 3-6 and 9-10 (Hap related models) for both happiness measures.

Regression

(21)

21

(3) Dependent variable III III III III

Coefficient of main regressor Hap1/2 and corresponding P-value -.157 (.419) (.972) .022 (.472) .157 (.634) .281 White-test for heteroscedasticity .025** .025** Instruments Female TSC TSC 1st_{stage F-statistic} _13.22 _38.52

Over-id test (P-value) .271 -

Endogeneity test

Woolridge .758 .821

(Adjusted) R-squared* .39 .38 .39 .39

(4) Dependent variable Hap1 Hap1 Hap2 Hap2

Coefficient of main regressor III and corresponding P-value

-.040

(.044)** (.415) -.068 (.620) -.008 (.118) -.097 White-test for

heteroscedasticity .000*** .209

Instruments Age Age Female

1st_{stage F-statistic} _12.85 _12.41

Over-id test (P-value) - .610

Endogeneity test

Woolridge .732 .115

(5) Dependent variable TSC TSC TSC TSC

Coefficient of main regressor Hap1/2 and corresponding P-value

.341

(.000)*** (.057)* .416 (.000)*** .568 (.126) .397 White-test for

heteroscedasticity .016** .064*

Instruments Female RI RI Whours

Over-id test (P-value) .790 .263

Endogeneity test

Woolridge .723 .469

Coefficient of main regressor TSC and corresponding P-value -.867 (.000)*** (.089)* .182 (.000)*** .200 (.054)* .192 White-test for heteroscedasticity .000*** .077*

Instruments Addict Age Addict

Female

Endogeneity test

Woolridge .982 .931

(22)

22 Coefficient of main

regressor Hap1/2 and corresponding P-value

.265

(.000)*** (.363) .127 (.000)*** .279 (.324) .135 White-test for

heteroscedasticity .000*** .003**

Instruments TSC Whours TSC Age

Endogeneity test

Woolridge .334 .261

Coefficient of main regressor RI and corresponding P-value .487 (.000)*** (.172) .276 (.000)*** .365 (.040)* .293 White-test for heteroscedasticity .000*** .077*

Instruments III III

Over-id test (P-value) - -

Endogeneity test

Woolridge .291 .602

Notes: 1) The sample size is 243

2) * significant at 10%; ** significant at 5%; *** significant at 1%. 3) Between parentheses are P-values belonging to the main regressors.

4) Where the over-id could not be performed due to a lack of instruments, a test was done to see if the instrument correlated with the error-term of the IV-regression. All results were 1.000 in P-values, meaning there was no correlation whatsoever.

5) Adjusted R-squared does not mean anything when used in IV-estimation so for IV estimation regular R-squared is reported. For OLS-estimation the adjusted version is reported unless there was heteroscedasticity reported, because then robust standard errors are used and with those, the adjusted R-squared becomes meaningless as well.

Before we begin comparing, notice that in all the models, none of the

endogeneity tests came back with a p-value below .10. This result means that we should trust in the OLS-estimates, because using IV when the regressors are

actually exogenous can be costly in terms of precision. These results are surprising though. Look for instance at regressions 5 and 6, which show very high scores for the endogeneity tests, meaning there is no endogeneity. But both regressions also score highly significant results for the coefficients, which are not close to zero. This is conflicting because the significant coefficients flow both ways, hinting that causality flows both ways, creating simultaneity bias, which is one of the main

(23)

23 causes of endogeneity, which we don’t find here. The methods used in this paper are not equipped to solve this problem so for now we assume there is no

endogeneity and look at the OLS-estimates.

By comparing regressions 3 and 4 to regressions 9 and 10, we see the difference between using III and using RI. If you look at the p-values of the

regression coefficients of the main regressors, you see the significance levels given by asterisks. Regressions 3-4 only have 1 statistically significant coefficient, for regression 4 at the 5% level, which means that the coefficient is true with 95% chance. In the context of figure 3, this means that only arrow 4 is true, and then only when using the Hap1 measure, because it is not significant in the model using Hap2. Regressions 9-10 however, find significant coefficients for all OLS models, using both happiness measures, and all at the 1% level. This shows that RI gets more meaningful results than III while predicting happiness.

Comparing Hap1, the absolute happiness measure, to Hap2, the more relative measure, by looking at significance done by comparing the left side of table 4 to the right side of table 4 shows only a slight difference. Hap1 is significant at the 5% level for regression 4, but not so for Hap2. In the other models, the measures perform about the same. One thing that does stand out is that III performs slightly better with Hap1, which can be seen by counting how many statistically

significant results are found in regressions 3 and 4. For Hap1, III only scores one result significant at the 5% level and for Hap2, III produces no significant results. Another result concerning the two happiness measures, Hap2 seems to explain about the same portion of the variance of the models as Hap1. This can be seen in that the R-squared statistics differ only slightly.

About causality little can be proven, but still these results suggest there could be causality where statistically significant results were found and vice versa. To put it in the context of figure 3, arrow 3 disappears, as there was no evidence

(24)

24 suggesting that happiness influences absolute income. Arrow four remains for Hap1 but not for Hap2 and arrows 5, 6, 9 and 10 remain as well.

4.2 Differences between III/ RI and Hap1/Hap2 for models on TSC.

This section takes the remaining regressions 1, 2, 7 and 8, which concern TSC and the income measures, and does the same as in the previous section. Table 5 shows those regressions using the same layout as table 4.

Table 5: OLS and IV-estimates on regressions 1-2 and 7-8 (TSC related models) for both happiness measures.

Regression OLS-Hap1 IV-Hap1 OLS-Hap2 IV-Hap2

(1) Dependent variable TSC TSC TSC TSC

Coefficient of main regressor III and corresponding P-value

.045

(.071)* (.228) .053 (.282) .025 (.522) .026 White-test for

heteroscedasticity .023** .078*

Instruments RI Whours RI Whours

Endogeneity test

Woolridge .819 .976

(2) Dependent variable III III III III

Coefficient of main regressor TSC and corresponding P-value .061 (.639) (.103) -.567 (.269) .146 (.357) .332 White-test for heteroscedasticity .028** .159 Instruments Hap1 Addict Hap2 1st_{stage F-statistic} _24.51 _38.52

Over-id test (P-value) .580 -

Endogeneity test Woolridge .028** .584 (Adjusted) R-squared* .39 .32 .38 .39 (7) Dependent variable TSC TSC TSC TSC Coefficient of main regressor RI and corresponding P-value .026 (.849) (.064)* .497 (.621) -.056 (.272) .311 White-test for heteroscedasticity .000*** .077*

Instruments III III

(25)

25

Over-id test (P-value) - -

Endogeneity test Woolridge .054* .173 (Adjusted) R-squared* .23 .18 .29 .26 (8) Dependent variable RI RI RI RI Coefficient of main regressor TSC and corresponding P-value -.028 (.417) (.352) -.072 (.256) -.036 (.555) -.044 White-test for heteroscedasticity .000*** .027**

Instruments Addict Age Addict Age

Endogeneity test

Woolridge .555 .912

2) * significant at 10%; ** significant at 5%; *** significant at 1%. 3) Between parentheses are P-values belonging to the main regressors.

4) Where the over-id could not be performed due to a lack of instruments, a test was done to see if the instrument correlated with the error-term of the IV-regression. All results were 1.000 in P-values, meaning there was no correlation whatsoever.

5) Adjusted R-squared does not mean anything when used in IV-estimation so for IV estimation regular R-squared is reported. For OLS-estimation the adjusted version is reported unless there was heteroscedasticity reported, because then robust standard errors are used and with those, the adjusted R-squared becomes meaningless as well.

Unlike the regressions in table 4, here we do find endogeneity, in the Hap1

regressions of models 2 and 7. This means that for those regressions we look at the IV-estimates instead of the OLS-estimates.

Comparing regressions 1 and 2 to 7 and 8 gives us the difference between III and RI. The results aren’t all that significant but unlike in the previous section they are equally so for III and RI. III finds one slightly significant results on the IV-estimates side, which is the side we look at due to exogeneity. RI also finds one, and like the result for III, both are only significant at a 10% level, meaning there is a 10% chance these coefficients are wrong. This was expected though, since the relation between self-control and income is not as widely recognized as the relation between income and happiness.

Only the models featuring Hap1 get any significant results, indicating that Hap1 is a better predictor in models with income and TSC. This seems logical,

(26)

26 since people with high self-control can only ever influence their own income, and not that of others, which is more a part of Hap2 since it compares happiness to that of other people. On another note, the Hap2 measure seems to have higher R-squared statistics overall, but this might be explained by noting that Hap2 has a higher correlation with TSC (and III) compared to Hap1, which basically says that more of the data concerning TSC is likely to be explained by Hap2.

Throughout the results section, we find that Hap1 has more statistically significant results compared to Hap2 and that they both seem to explain the data about equally well. Still, Hap2 shows to have equally high correlations with the income measures and TSC compared to Hap1. Hap2, which compares happiness to others could very well be a more descriptive statistic than just absolute

happiness, and it does not do much worse than Hap1 in the models. Perhaps an aggregate measure using both could replace the purely absolute measure that is still most widely used. The exact relationship should be examined more closely in the future, like Alpizar et al. (2005) did. Maybe coming up with a new set of questions altogether, one that mixes the absolute and the relative perspectives.

The results of the regressions concerning TSC suggest that in the context of figure 3 causal relations, or arrows, 2 and 8 do not exist. Arrows 1 and 7 both do exist, but only at a 10% statistical significance level for the Hap1 regressions. Combined with the results from the previous section, these results suggest there only exist causal relations as shown in figure 4 below.

(27)

27 It must be said that these are just suggested causal relations, there is no proof confirming this. Still, this is a confusing result, throughout the literature, AI, represented here by III, has mostly shown very significant results predicting happiness, and often reverse causality. This was not found in this research, instead that result is found with RI. It might be that there were simply too few observations to make statistically significant results in the regressions concerning III and happiness. What is does show however, is that RI seems to predict

happiness better than III.

This could mean RI, relative income, is a better predictor than AI, absolute income, but it does not have to be so. III, the AI measure used in this paper, has its limitations due to it being a reported measurement. Relying on self-reported income instead of a more conclusive income measure means inviting biases. This is a clear limitation of the data.

On the other hand, RI is also a self-reported measure, inviting the same biases. Considering that, the RI measure still predicted better than the AI measure. This result validates the case made by happiness economics. Regardless of the

restrictions of the data caused by using self-reported subjective measures, this result is evidence suggesting that RI is at least as good a measure as AI regarding the relationship with utility/happiness, if not better, and that RI could replace, or help replace, AI as a measure of utility.

4.3 Mediation analysis

To test whether there is mediation, or rather to identify the underlying process regarding happiness and income, a simple mediation analysis was done. Here, according to the steps laid down by Baron and Kenny (1986), it was tested if the predictive power of AI, here III, came from a mediator, i.e. RI. This requires 4 steps, which will be explained using figure 5 below.

(28)

28 Step 1 is finding the correlation depicted by C, step 2 is finding the correlation depicted by A and step 3 is finding the correlation depicted by B. Then lastly, step 4 means regressing the dependent variable on both the mediator and independent variable to confirm that the mediator significantly predicts the dependent

variable. This gets us a new ‘mediated’ correlation C’ which should be reduced. If all these steps are met, then the mediator completely mediates the relationship known as C, making C meaningless. The results of those steps can be seen in table 6 below.

Table 6: Mediation analysis for Hap1 and Hap2 per step

Regression Step1 Step2 Step3 Step4 Coefficient

Mediator RI Hap1 Coefficient of III

and corresponding P-value .1224 (.523) (.000)*** 1.787 (.000)*** .472 (0.053)* -.038 (.000)*** .539 (Adjusted) R-squared* .00 .16 .13 .14

Hap2 Coefficient III and corresponding P-value .0398 (.014)** (.000)*** 1.787 (.000)*** .405 (.878) .003 (.000)*** .401 (Adjusted) R-squared* .02 .16 .13 .13

2) * significant at 10%; ** significant at 5%; *** significant at 1%. 3) Between parentheses are P-values belonging to the regressors. 4) Mediator coefficients of variable RI belong in step 4.

The mediation analysis using Hap1 fails immediately as III fails step 1, it does not significantly influence Hap1. This is quite surprising given the literature. For the other mediation analysis, there is no such problem. All steps were completed and the requirements met, all the necessary relations were significant. This means that

(29)

29 RI fully mediates III. This is partly proof that most of the literature is wrong

because RI instead of AI seems to be predicting Happiness, making the causality assumed in most literature faulty. It is also further evidence that different utility measures give different results. Highlighting the need for a new and better utility measure that considers the need to compare to others.

5. Discussion and Conclusion

The Easterlin paradox showed us that absolute income does not necessarily have such an impact on happiness, and recent research by Deaton and Stone (2013) tells us that relative income does, and that replacing AI with RI solves the Easterlin paradox. This is in line with the findings of this paper. Here however, the results also indicate that RI has as much a relationship as AI with at least one of the determinants of happiness, namely trait self-control, or TSC, which was particularly visible in figure 4. Suggesting that the answer to the question if relative income can help replace absolute income is yes.

Testing the two happiness measures provided another piece of the puzzle. Expected was that Hap2 was correlated more to RI than to III, the AI measure, since RI and Hap2 were both relative. This was true, since table 4 showed that Hap2 regression results were only significant for RI and not for III. Furthermore, III did only a little better for Hap1 than for Hap2 in table 4. The amount of

variance explained, represented by the adjusted R-squared statistic, was also similar for both happiness measures. If you recall, Hap2 was the more relative happiness measure and Hap1 the more absolute happiness measure. These

findings suggest that using Hap2 is about as good as using Hap1 when modelling with income measures.

After finding that RI did better than AI, a mediation analysis was done to see if perhaps the classic model which uses AI as a predictor of happiness was mediated by RI. The results of this analysis did indeed show that RI fully

(30)

30 mediated the relationship between III and happiness. This indicates that the classic model is wrong and that in fact relative income determines happiness.

The results that using relative happiness works as good as absolute happiness and that RI does better than, and in fact fully mediates, AI, seems to indicate that people, who are assumed to be rational agents, compare a lot more than assumed in economic theory. This finding probably would not come as a surprise to a neuro-economist or someone who has studied the psychology involved, but it does indicate that traditional assumptions used by economists are not correct. How then should we use this ‘new’ knowledge? The results from regressions 3, 4, 9 and 10 suggest that both the income measures influence happiness, unequally so, maybe we should do as Alpizar et al. (2005) attempted by finding how much each influence happiness and use that to construct a new income measure, incorporating what has been learned. We know that AI and RI both influence happiness, so why not make a new measure that combines the two?

By the same logic, the findings on the happiness measures show the need for a new comprehensive utility measure. There are many currently used in literature such as well-being, satisfaction with life, and general happiness measures. Most of these being rather ‘loosely’ defined, by which I mean that researchers pick a measure of utility and stick a name to it such as well-being, making assumptions about their effectiveness based on other research. This is not a terrible thing, being a widely accepted practice in research, but nowadays there is maybe too much choice. Especially when the goal is to get economists to accept innovative ways. This is a concern which can maybe be solved by Neuro-economics, which looks to the brain to confirm or reject hypotheses and assumptions. This way, they

confirm or deny ideas that economists have and limit the options so to speak. Limitations of the research

The use of a questionnaire might have limited the research in multiple ways, most notably lack of control and a lack of validity. There is also the issue of

(31)

31 subjectivity in answering the questions used and some level of researcher

imposition, meaning that the researcher decides beforehand what information is important and what is not, which makes it possible to miss vital information. Of course, these issues can be anticipated and even negated a bit, but it is unlikely to fully counter these problems in any given setting.

Furthermore, the data was acquired via two channels, Facebook and Mturk. As discussed before this should not have caused much problems but it would have been better to use just one channel, particularly because the motivation would have been more equal for all subjects.

There is also the number of observations. After omitting subjects for assorted reasons, only 243 subjects remained. This limited dataset invites weaker statistical significance in the models.

Suggestions for further research

There are a lot of avenues for future research, and some are already being investigated. Still, until they are fully explored they are worth mentioning again. The models used in everything related to happiness economics must all account for causality and the problems that accompany this. This has been studied before and some few conclusions have been drawn, but the results have not been

conclusive and sometimes even contradictive. Either way, this clearly remains a topic of future research.

Another topic is about the use of subjective measures, here we use subjective relative income which was used by Mangyo and Park in 2011 and was proven useful. Such a subjective measure seems to work just as good as absolute measures and is probably cheaper to measure. And just as Lindner et al. (2015) did with the brief self-control scale, which used 13 questions instead of 30 and worked just as good as the full version, we need someone to test and prove whether this subjective measure works just as good as the absolute measure.

(32)

32 But differences in methods have the same effect. Like Brown et al. (2015) shows us, we still do not know if effects of relative income on utility are negative or positive. As they show in their paper, some methods say positive, some say negative, this is hard to accept from an empirical standpoint. Maybe some additional research can be done to find a ‘best practice’, or we could just follow the advice Brown et al. give us, which is to make a convincing case for the methodological choices made and to prove the robustness of the results. Conclusion

Recent literature found that relative income ‘solved’ the Easterlin paradox. This paper adds to the literature by ‘testing’ relative income in established settings, bringing two new contributions to that literature. Firstly, RI better predicts

happiness than AI does and does equally well predicting one of the determinants of happiness, namely TSC. If this holds for other determinants it would be proof for using RI and proof against using AI. Secondly, the mediation analysis finds that RI fully mediates the relation between AI and happiness, which is definitive proof that RI should be used instead of AI.

Two different measures of happiness, one absolute and the other more relative, both did well as a proxy of utility. The absolute measure has been used traditionally and that seems to be a good thing, since it got more significant results in this research. However, the relative measure seemed a good substitute, which suggests that utility is not only derived by absolutes and that subjects compare more than assumed in economic theory.

Can relative income help replace absolute income? And to what degree? The results argued for a bit of both, since both RI and AI seem to influence happiness, with RI significantly more so. The results of this paper advocate making a new income measure that uses both, maybe using proportions based on research such as Alpizar et al. (2015) used. Also, if choosing one or the other happiness measure changes results, which it does in this paper, it might be logical to do the same for

(33)

33 a new happiness measure.

Bibliography:

Alpizar, F., Carlsson, F., Johansson-Stenman, O. (2005). How Much Do We Care About Absolute Versus Relative Income and Consumption? Journal of Economic

Behavior & Organization. Vol. 56, 405-421.

Baron, R. M., & Kenny, D. A. (1986). The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical

considerations. Journal of personality and social psychology, 51(6), 1173.

Brown, S., Gray, D., Roberts, J. (2015). The relative income hypothesis: a comparison of methods. Economic Letters. Vol 130, 47-50.

Cheung, T. T., Gillebaart, M., Kroese, F., & De Ridder, D. (2014). Why are people with high self-control happier? The effect of trait self-control on happiness as mediated by regulatory focus. Frontiers in psychology, 5.

Deaton, A., & Stone, A. A. (2013). Two happiness puzzles. The American economic

review, 103(3), 591-597.

De Neve, J. E., & Oswald, A. J. (2012). Estimating the influence of life satisfaction and positive affect on later income using sibling fixed effects. Proceedings of the National

Academy of Sciences, 109(49), 19953-19958.

Diener, E., Emmons, R. A., Larsen, R. J., & Griffin, S. (1985). The Satisfaction with Life Scale. Journal of Personality Assessment, 49(1), 71-75.

Diener, E., Kesebir, P. (2013). A Virtuous Cycle: The Relationship Between Happiness and Virtue. Retrieved at

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2309566

Di Tella, R., & MacCulloch, R. (2008). Gross national happiness as an answer to the Easterlin Paradox? Journal of Development Economics, 86(1), 22-42.

Easterlin, R.A. (1995). Will raising the incomes of all increase the happiness of all?

Journal of Economic Behavior and Organization, 27(1), 35-47.

Easterlin, R.A. (2001). Income and Happiness: Towards a Unified Theory. The

Economic Journal, 111(473), 465–484.

Ferrer-i-Carbonell, A., (2002). Income and Well-being: An Empirical Analysis of the Comparison Income Effect. Journal of Public Economics, 89(5), 997-1019.

(34)

34 FitzGerald, T. H., Friston, K. J., & Dolan, R. J. (2012). Action-specific value signals in reward-related regions of the human brain. Journal of Neuroscience, 32(46), 16417-16423. http://doi.org/10.1523/JNEUROSCI.3254-12.2012

Formann, A. K. (1984). Die latent-class-analyse: Einführung in Theorie und Anwendung. Beltz.

Fowler, S., Willis, G., Moser, R., Ferrer, R., & Berrigan, D. (2015). Use of Amazon Mturk online marketplace for questionnaire testing and experimental analysis of survey features. In Proceedings of the 2015 Federal Committee on Statistical Methodology

Research Conference, Washington, DC. Retrieved from http://sites. usa.

gov/fcsm/files/2016/03/C2_Fowler_2015FCSM. pdf.

Frey, B. S. (2008). Happiness: A revolution in economics. MIT Press Books, 1. Retrieved from http://www.jstor.org/stable/j.ctt5hhhtv.6

Hajek, A. (2013). Endogeneity in the relation between poverty, wealth and life satisfaction. The German Socio-Economic Panel Study (SOEP) at DIW Berlin, 604. Hofmann, W., Luhmann, M., Fisher, R. R., Vohs, K. D., & Baumeister, R. F. (2014). Yes, but are they happy? Effects of trait self‐control on affective well‐being and life satisfaction. Journal of Personality, 82(4), 265-277.

International Monetary Fund. (2017) World Economic Outlook. Retrieved from https://www.imf.org/external/pubs/ft/weo/2017/01/weodata/index.aspx

Li, H., Liu, P. W., Ye, M., & Zhang, J. (2011). Does money buy happiness? Evidence from twins in urban China. Retrieved from

http://scholar.harvard.edu/files/maoliangye/files/ye_march6_2014_doesmoneybuyha ppiness.pdf?m=1394431202

Lindner, C., Nagy, G., & Retelsdorf, J. (2015). The dimensionality of the Brief Self-Control Scale—An evaluation of unidimensional and multidimensional

applications. Personality and Individual Differences, 86, 465-473. DOI: 10.1016/j.paid.2015.07.006

Lyubomirsky, S., & Lepper, H. S. (1999). A measure of subjective happiness:

Preliminary reliability and construct validation. Social indicators research, 46(2), 137-155.

Mangyo, E., Park, A., (2011). Relative Deprivation and Health: Which Reference Groups Matter? Journal of Human Resources. 46(3), 459-481.

McBride, M. (2001). Relative-income effects on subjective well-being in the cross-section. Journal of Economic Behavior & Organization, 45(3), 251-278.

(35)

35 Moffitt, T. E., Arseneault, L., Belsky, D., Dickson, N., Hancox, R. J., Harrington, H., ... & Sears, M. R. (2011). A gradient of childhood self-control predicts health, wealth, and public safety. Proceedings of the National Academy of Sciences, 108(7), 2693-2698. Oswald, A. J., Proto, E., & Sgroi, D. (2015). Happiness and productivity. Journal of

Labor Economics, 33(4), 789-822.

Raman, B., & Chebrolu, K. (2007). Experiences in using WiFi for rural internet in India. IEEE Communications Magazine, 45(1), 104-110.

Sarstedt, M., & Mooi, E. (2014). A concise guide to market research. The Process, Data,

and Methods Using IBM SPSS Statistics.

Stevenson, B., & Wolfers, J. (2008). Economic growth and subjective well-being:

reassessing the Easterlin paradox (No. w14282). National Bureau of Economic Research.

Stutzer, A., & Frey, B. S. (2006). What happiness research can tell us about self-control problems and utility misprediction.

Yu, Z., & Chen, L. (2016). Income and well-being: Relative income and absolute income weaken negative emotion, but only relative income improves positive emotion. Frontiers in psychology, 7.

Appendix A:

Explaining the Vetting Process:

At the start there were 423 subjects that filled in the survey.

Wanted subjects were between 20 and 65 years old, were enjoying an income and filled in their questionnaire honestly. 69 Subjects were excluded because of this. People outside the mentioned age limits were therefore excluded instantly (37 Subjects). People who did reported an income of 0 were excluded (8 Subjects). So, non-workers who reported an income were not excluded.

Furthermore, I had to exclude some subjects who misunderstood my remark about rounding their income to thousands (8 Subjects). Either they filled in numbers that did not seem possible given their working hours or they did not fill in anything meaningful, like a dot, a zero or an exclamation mark.

For the same question subjects also converted to dollars when living outside of the USA, these were also excluded because there was no way to undo it without knowing their used exchange method (5 subjects).

Lastly, some workers of Mturk tend to try to cheat, which was why I entered 2 questions to detect this. One said: please select Very much. The other measured the time the subjects took to fill in the form. I excluded all workers that did not answer with Very much and all workers that filled in the entire questionnaire under 2

(36)

36 minutes (11 Subjects). Simply because they did not or could not answer truthfully. This leaves 354 subjects. Then it was decided to only use data from the Netherlands and the USA. This left 243 subjects that comprise the dataset of this paper.

The Full models per table:

As an example look at regression 1 of table 1. That would essentially mean this model is used for OLS. 𝑇𝑆𝐶 = 𝑎1+ 𝛽1𝐼𝐼𝐼 + 𝛽2𝐻𝑎𝑝2 + 𝛽3𝐴𝑔𝑒 + 𝜀1

Where III is instrumented by GI and Whours in the case of 2SLS. Table A1: Models of regressions 1-10 for two happiness measures Regression Dependent

variable Main Regressor Additional regressors Instruments 1 TSC III Addict, Age & Hap1 RI & Whours 1 TSC III Addict, Age & Hap2 RI & Whours 2 III TSC Whours, RI & Age Hap1 & Addict

2 III TSC Whours, RI, Age &

Addict Hap2

3 III Hap1 Whours, Age & RI Female & TSC

3 III Hap2 Whours, Age & RI TSC

4 Hap1 III RI, TSC & Female Age

4 Hap2 III RI & TSC Age & Female

5 TSC Hap1 Addict & Age Female & RI

5 TSC Hap2 Addict & Age Whours & RI

6 Hap1 TSC RI & Whours Addict & Age

6 Hap2 TSC RI Addict & Female

7 TSC RI Age, Addict & Hap1 III

7 TSC RI Age, Addict & Hap2 III

8 RI TSC Hap1 & III Addict & Age

8 RI TSC Hap2 & III Addict & Age

9 RI Hap1 III TSC & Whours

9 RI Hap2 III TSC & Age

10 Hap1 RI TSC, Female & Whours III

10 Hap2 RI TSC III

Notes: 1) The first of two of the same regressions always uses Hap1 and the second always uses Hap2.

2) Data from the Netherlands and the USA, which makes for 243 subjects. 3) These are the models for tables 3 and 4 in the results section.

Appendix B: Questionnaire:

The questionnaire existed of a small block 1, questions 1-7, and a big block, block 2, that came after, which was randomized. The big block houses both the happiness scales and the TSC scale. How the scales work is explained for each of the measures separately. All notes are in italics.

Block 1:

(37)

37 q2: What is your age?

q3: How many hours a week do you work on average? q4: Please evaluate yourself for the following questions.

From left to right, 1 to 5 points. Add them for the addiction measure used as an instrument.

q5: What would you say your gross income is (rounded to thousands)? If you don’t know, just make an educated guess. Also, please denote the currency if it’s not in dollars.

q6: Compare your average living standard (income) with the groups below, do you feel your living standard is much better, much worse or anything in between?

From left to right, 1 to 5 points. Add them for the RI measure as described in Mangyo & Park (2011).

q7: Do you think a lot about your income and how to improve it?

Block 2:

(38)

38

Question 8 constitutes the first happiness measure, the Satisfaction with Life scale.

From left to right, 7 to 1 points. Add them for the Hap1 measure as described in Lyubomirsky & Lepper (1999).

Question 9-12 constitute the second happiness measure, the General Happiness Scale.

From left to right, 1 to 7 points. Add them for questions 9, 10 and 12. Then invert question 11 and add that for the Hap2 measure as described in Diener et al. (1985).

q9: Please pick the options that you think describe you best.

q10: Please pick the options that you think describe you best.

(39)

39 q12: Please pick the options that you think describe you best.

q13: Do you think a lot about happiness and how to improve it?

Question 14-16 constitute the Brief Self-Control Scale with 13 questions.

Invert questions 2-8, 10-11 and then add the points to get the TSC measure as described in Lindner et al. (2015).

q14: Please indicate how much of each of the following statements reflect how you typically are.