Comparison Income Effect on Subjective Well-Being Abstract We follow the comparison income effect study on subjective well-being in Ferrer-i- Carbonell (2005), and test the robustness of those results by using longer time period, different income measures and a two-stage estimation method to estimate the reference income. We also check the convergence of the income comparison pattern in the East and West Germany. Meanwhile, as an extension of the restudy, we introduce some other ways of defining the reference group in the study of comparison income effect and also we implement a first differenced ordered logit model with fixed effects instead of the ordered probit model with random effects. We show that the causality of the comparison income effect is coming from the different ranking level of an individual within the same reference group. Keywords: Comparison income; Interdependence of preferences; Reference group; Relative utility; Subjective Well-being; Rank and inequality; Transitional economy.
1. Introduction The utility theory has been studied intensively for about two centuries and still has many unknown propositions left for us to explore. The importance of the utility study is that the maximization of the individual utility level has been used as the decision rule of a rational person. And for the policy maker or the central planner, in order to maximize the social welfare, he also needs to refer to the characteristics of the social welfare function which is a combination of all individuals utility levels in the whole society. In light of the utility theory study, we can not overlook the impact of income on the utility estimation. If an individual is insatiable, we believe that higher income would definitely bring this individual higher utility level since an individual can buy more consumption through higher income. So the relationship between income and utility level is the most important and also a very complicated part in the utility theory. In the past two decades, many studies of the Subjective Well-Being (SWB) have been worked through by the psychologists and the sociologists. Diener, Suh, Lucas, and Smith (1999) define the subjective well-being and summary the use of it in social science. Headey, Veenhoven and Wearing (1991) and van Praag, Frijters and Ferrer-i-Carbonell, 2000) check the causality between the subjective well-being and some objective measurements. They contribute a lot in the behavior theory. Further Frey and Stutzer (2002) try to implement these achievements into economic study and explore the ways in which we can relate the income impact with the level of SWB. 1
Unlike the utility theory in economics which is derived as a function of consumption levels, the SWB is a measure of self-reported happiness. Some psychologists have shown that the income level only has impact on an individual s SWB when his income is below some threshold. Here the income level enters the welfare function as an absolute value and it is different from the way we use the levels of consumption which are constrained on a certain level of income to measure an individual s utility level. Some research has shown that the higher income level can not always bring people higher happiness or feeling more satisfied with their life. This means the positive correlation between income and SWB may not be so strong as we thought. Start from this fact, it is found that SWB depends not only on the absolute value of an individual s income level, but also depends on some other specific properties of this individual s SWB function. Some of the properties are coming from this individual s own experiences, such as the living environment, education level, occupation, family composition, etc. And some of the specific properties are coming from the other individual s influence in the same society. Here this kind of situation is referred as the interdependence of preferences. And this comparison income effect could be helpful in order to explain the reason that higher income of each person in the same society may not make all people feel happier (see, e.g., Easterlin, 1995; Kapteyn, van de Geer, Van de Stadt, 1997). When the relative income remain unchanged for each individual in the society, we may observe no change of the SWB if we believe that people are used to compare their life 2
situations with those of the ones who live with them in the same society. By the same society, we mean a group of people with similar individual characteristics. Here we can also call this type of society as a reference group. Each individual has his own reference group, and he measures his SWB by valuing his relative performance in his own reference group. If we can find a way to define a good reference group for each individual, we can measure the impact of comparison income or the reference group s mean income on an individual s SWB besides the effect coming from his absolute income. Following the same method of defining the reference group in Ferrer-i-Carbonell (2005), we categorize individual s characteristics into several subgroups and form a reference group by choosing a unique combination of these category levels. In this paper, we try difference understandings of the welfare function and use the empirical results to show how much the reference group s income level contributes to people s SWB. Comparing with the previous research, we try some new ways to examine the comparison income effect. Firstly, we follow the specifications in Ferrer-i-Carbonell (2005) to test the robustness of her results by using a panel data set with longer period and introduce the individual labor income into the welfare function with some different ways of defining the reference group. Secondly, we include the rank levels of individuals in the same reference group and find out that there is not only effect from the absolute value of the reference income, but also effect from the relative position of an individual in his reference group. Thirdly, we study the convergence of the East to the West on the 3
comparison income effect based on the revised model we get after we include the ranking effect. The paper is structured as follows. Section 2 reviews the literature on the comparison income effect and the relative utility. Section 3 explains the steps of problem solving in details. Section 4 introduces the data we used for empirical analysis and explains the way we construct the model empirically. Section 5 contains the empirical results from section 4 and the last section concludes the findings and also suggests some further research directions. 2. Literature Review People live their life not only as an independent individual, but also as a member of a group or community. And the behavior of other people in the same society could affect the decisions made by this individual. Related researches have been done also in the consumer studies. The herding behavior studied by sociologies has pointed out that people use to make their own decision based on others decisions on the same issue (see, e.g., Banerjee, 1992). In Akerlof and Kranton (2000), they argue that individuals picture themselves through a group of social economic standards and fit themselves in a group of people who have the similar levels of standards and they call it economic identity. In this paper, we take the assumption that the reference group s income has effect on people s SWB and we use the empirical analysis to verify this hypothesis. To find out in 4
which way the others income levels affect an individual s SWB, we study the effects from two different reference income measures. One is the average income level of the reference group, and the other is the relative rank of this individual in the same reference group. In the empirical result, we will show which measure is more effective as the comparison income effect. Paper by Oswald and Clarke (1996) is one of the paper that relate the comparison income and satisfaction level together. Our research here is based on their empirical founding and trys to check how much the definition of the reference group matters. The absolute level of reference group gives us the information about how the average guy in this reference group performs. And conditional on an individual s income, we expect negative impact from a higher reference income on this individual s SWB. The intuition here is when an individual observes a better average performance of the other individuals in the same reference group, he may feel unhappy even he remains the same income level as before, because he may feel less successful by comparing with the others in the same reference group. In order to check the effect of the reference group income, we have to pick the right reference group base on individuals preferences. In this paper, we try several combinations of the reference group definition and check the different empirical effects from these different reference groups. We can perform some sensitive analysis based on these empirical results we get and get a better idea on how people are affected by the reference income. In some circumstance, people compare their life standards with others 5
life condition as a member of family. In the other circumstance, an individual evaluates his social position based on both of his and others individual characteristics. In this paper, we construct two kinds of reference income, one is the reference family income and the other is the reference individual labor income. We try to find out which reference income is more appropriate in our welfare function in order to get more reasonable reference income effect. The theoretical foundation behind this empirical method is the unitary and collective model for household resource allocation. The rank level is also important in the study of the comparison income effect. Not only the absolute value of the reference income matters, the distribution of the reference income can also bring us some information on how people value their income referring to the others income levels in the same reference group. Some works have been done by checking the impact from different inequality indices such as GINI index, for example, Clark (2003). It is shown that higher inequality level may inform the individuals that there are more opportunities to get higher income levels in the future and people in this reference group would expect their life time income in a more optimistic way then report higher SWB levels. In our paper, instead of using the macro inequality measures, we use the cumulative distribution function (CDF) s level for each individual to check the impact of ranking position on his SWB measurement. We expect a positive relationship between the CDF level and the SWB level since higher ranking indicates that an individual has higher social status and this would make him feel happier compare with the others with lower economic position. 6
From our results in this paper, we find out that there is significant comparison income effect when an individual measures his SWB. Based on different combinations of reference group, we get significant positive effect from the ranking of an individual in his reference group, but only some significant negative effect from the absolute value of the reference income. It shows that people care more about the relative position in the reference group than the aggregate condition of their reference group. 3. Model 3.1. SWB measurement The measurement we use in this paper is the general satisfaction of life which is the level of satisfaction an individual grades for his current life in a general way. This measurement is a self-reported measure of SWB, so that we don t need to predict the individual s utility level through some complicated estimation procedures. The question for the SWB measurement in the survey is the following one: How happy are you at present with your life as a whole? The answer of this question has 11 different choices, in which 0 means totally unhappy and 10 means totally happy. 7
The use of this kind of self-reported life satisfaction measures are studied in a great extent (see, e.g., Ferrer-i-Carbonell, 2002). Due to the personal heterogeneity of the reported life satisfaction level, we have to make some assumptions before we bring the survey data into the empirical analysis. Detailed discussion is in the estimation method subsection. 3.2. Welfare Function The welfare function we refer here is a measure of an individual s well-being. In this paper, we implement a subjective measurement which is the general satisfaction of life to study the impact of income on level of well-being. Unlike the tradition way of measuring the individual utility level by using the bundle of consumption and leisure, we treat welfare as a function of income and some individual characteristics. The reason of this refinement is that the income levels are easier to measure and also easier to compare than the measures of consumption and leisure in the utility function. To test the hypothesis that there is a significant relationship between the reference income and an individual s SWB, we start from the format of the individual welfare function: SWB = U( Y, Yr, X, SWB 1), (1) 8
Here we use an individual s income vector Y, his reference income vector Y r, a group of his individual characteristics X and his last period s SWB level to estimate his current level of SWB. Here the individual income vector includes the individual labor income and the partner s income. The reference income vector includes the reference individual income and the reference family income. We also add the cumulative distribution function of the individual and family income rank into the reference income vector. The individual characteristics include gender, age, marital status, number of children in household, number of adults in household, status of employment, years of education, and region. The reason we put lagged SWB level in the welfare function is that the unit root test result with a P-value of 0.4229 tells us that the last period s SWB does have effect on the current level of SWB. The contribution of this paper on the relationship between income and well-being is on the new way of including income covariates into the welfare function. We include both of the individual income and the family income in the study of income impact on well-being level. For the absolute income vector, we use the individual labor income and the partner s income instead of using the family income directly in the welfare function as the way studied in the previous research. The advantage of using both of the individual labor income and the partner s income instead of the family income as a whole is through these two income categories, we can check the validity of the unitary and collective model of intra-household resource allocation. According to the unitary model, a couple in 9
the same family would make their decisions based on the sharing of the total of their individual income and the more the happier for both of them. But for the collective model, there is higher bargaining power for the one with higher income and each individual may make his consumption choices mainly based on his own labor income level. In our empirical analysis, we use both of the individual labor income and the partner s income to check the absolute income effect and use the individual reference income and the family reference income to check the relative income effect. And when we construct the reference group, we fit several different combinations of reference group criterions and check whether the comparison income effect is robust to the composition of the reference group. One of the reference income definitions is the one used in Ferrer-i- Carbonell (2005) which is the average income for a group of people with the same value of age, years of education, gender and region. In addition to that, we add two more ways of defining the reference group. One is by adding the employment status into the above mentioned reference group. The other one is by adding the number of kids into the original reference group. The results from checking these two more different reference groups can tell us whether the composition of reference group affects the level of the relative income effect. We will see the contribution of these improvements in the empirical results section. In order to check the robustness of the results in Ferrer-i-Carbonell (2005), we follow the four specifications performed in Ferrer-i-Carbonell (2005) which are the following ones: 10
In the first specification, we only include the individual labor income and the partner s income in the welfare function. In this step, we can check both of the sigh and size of the absolute income effect on an individual s SWB. The second specification adds the reference incomes into the welfare function. We try several different compositions of reference group. In this case, we can check the effect from the absolute values of the reference income. Additional to the absolute values, we introduce the income ranks of an individual and his family in his reference group. We believe that the relative position of an individual in his reference group also matters and by checking this impact, we can see whether people care more about their relative position than the absolute average levels of their reference group. In the third specification, instead of using the reference income, we use the distance between the individual or household income level and the reference income level. Although the new distance covariate does not change the welfare function econometrically, on the economics point of view, the new covariates bring us some new intuition to explain the comparison income effect. Here, the sign and size of the coefficients for the distance covariates would show us how much people care about the income difference from an average guy. In the last specification, we use another two new covariates to check whether the effects from the income distances are symmetric for the poorer and richer guys comparing with 11
the average guy. This last specification may lead us to some policy issues, such as the tax and subsidy regulation in the case that there are asymmetric effects from the poorer and richer distances. In each of these four specifications we try several different reference group combinations. One is the same as the one used in Ferrer-i-Carbonell (2005). The others are the ones we use to check the sensitivity of the comparison effect. We add several more covariates to make the reference group more specific to individuals characteristics. The additional covariates are the employment status and the number of kids in the family. Based on these difference reference group compositions, we get different comparison income effects. We will discuss the differences in more details in section 5. Besides the four specifications in Ferrer-i-Carbonell (2005), we also implement the twostage estimation method in Senik (2003) to see whether we can get robust estimators by using different approaches. The first stage of the two-stage estimation is to estimate the predicted incomes from the individual characteristics and the second stage is to substitute the predicted incomes as the reference incomes into the welfare function and check whether there is any significant impact from the predicted income levels. By using this two-stage method, we avoid the trouble to define a suitable reference group. We will discuss the shortcoming of this method in section 5 following the empirical results of the four specifications as those discussed in Ferrer-i-Carbonell (2005). 4. Data and Estimation Method 12
4.1. GSOEP The data we use to perform the empirical analysis is a subset of the German Socio- Economic Panel (GSOEP). This data set collects the West German data before the Unification started from 1984 and both of the East and West German data after the Unification till 2002. Thanks to the high quality of this survey data, we can check the impact of this non-experimental event on the SWB before and after the Unification and also the convergence pattern of East and West Germany following the Unification. In this paper, we use a similar data set as the one used in Ferrer-i-Carbonell (2005) but we double the length of the time period. We use this updated data set to check the robustness of the results in Ferrer-i-Carbonell (2005). 4.2. Estimation Method The SWB levels we explore in this paper are coming from the individual self-reported general life satisfaction level. We need an ordered probit or ordered logit model to fit the latent general satisfaction levels rather than the reported levels directly. Before we decide whether we should use the probit or logit model, we need to test the existence of fixed effects, between effects and random effects. In order to deal with the between effects, we can simply add year dummies to check yearly differences. To decide whether we should choose fixed or random effects, we need to perform the Hausman test. 13
The empirical model we apply to the data set is the following one: SWB nt k = + β T + β ynt + β yr nt + β SWBnt + + 3 1 2, 3 1 β i xi, nt i= 4 α 0 + v + ε, (2) n nt where years, n T is the subscript for different individuals and t is the subscript for different is the year dummy vector to capture between effects, k is the number of covariates for individual characteristics, v n is the individual fixed or random effect, and ε nt is the random error term. In the above welfare function, we have a group of individual characteristics, a group of income levels, year dummies, lagged SWB level and error terms. We need to separate the error term into two parts in order to get more efficient estimators. One part is the random error term which is uncorrelated with any explanatory covariate. The other part is the fixed or random effect coming from individual heterogeneity. If we assume there is a random effect from each individual, we believe that the random effect is uncorrelated with any explanatory covariate, and it is normally distributed. By assuming that we have random effects here we make the estimation easier to perform. In the other way, we could assume that there is a fixed effect from each individual. In this case, we believe that some of the individual characteristics are highly correlated with the unobserved individual heterogeneity. Under this circumstance, we add more covariates 14
into the model fitting and we lose degrees of freedom. Since there are many unobserved individual characteristics included in v n and those may have impacts on the individual income levels, we should consider fixed effects in our estimation of welfare function even that we have to sacrifice degrees of freedom. These unobserved individual characteristics may not be normally distributed. The Hausman test s results with a 2 χ value is 141.21 and a P-value as 0.0000 reject the null hypothesis that random effects are as consistent as fixed effects. In this paper, we perform an ordered logit model with fixed effects in order to find out the comparison income effect on SWB. The nature of the panel data is not a benefit for us to estimate fixed effects directly. What we do in this paper is to use first differences estimation. By taking the first differences we can get rid of the fixed effects problem and leave a stationary process for our empirical analysis with P-value as 0.0373 from the unit root test. Equation (3) is the model we apply in the ordered logit model. Here we pick the logit model because the original model has fixed effects and probit model is more suitable for a model with random effects. k SWBnt = β T + β ynt + β yr nt + β SWBnt + + 3 1 2, 3 1 β i xi, nt i= 4 0 + ε (3) nt To make the estimation consistent, we need to use the second lagged SWB level as instrument variable of the first lagged differenced SWB. 15
Before we discuss the empirical results in the following section, we also need to check for endogeneity from the interaction between the income and the SWB levels. According to recent researches on the relationship between people s behavior and their satisfaction of life, it seems to be a mutually relationship. We know that higher income level will bring an individual higher satisfaction level. Meanwhile, happy people would be likely to participate more in the labor market and earn higher income. When we add the explanatory variables into the differencing equation, we get that the coefficient of the individual labor income is significant from zero (P value is 0.029), so we should construct the IV estimation of it in the first step and use the result into the second step which is the ordered logit model. Here we use the lagged income as instruments and perform the Hausman test again to confirm for the solving of endogeneity problem with a P-value of 0.068. 5. The Empirical Results 5.1 Basic Specifications In this section, we give the results from the different specifications mentioned in the third section. We explain the findings and compare them with the ones in Ferrer-i-Carbonell (2005). 16
In the first specification, we only use the individual income and the partner s income to check the effect of income on an individual s SWB. In Table 1 we see that as a full sample and also in the West, we get negative effect from the increased labor income, this may caused by the higher income expectation conditional on the increased income, but the impact from the partner s income is much lower than the one from the individual income and also insignificant. Since there is no significant effect from the partner s income, we see that separate the individual labor income and the partner s income in the regression is more preferable than using the family income instead. The F-test result (p=0.1414) shows that in East Germany the differenced measures have no effect on people s differenced satisfaction level. These results show that an individual would assign their individual labor income a higher weight than that for his partner s income. When we compare the results for East and West Germany, we see that the individual income effect of the East is lower then that of the West. But the effect from the partner s income in the East is higher than that in the West. Comparing these results with those in Ferrer-i-Carbonell (2005), we see that in Ferrer-i- Carbonell (2005), the family income effect in the East is higher then that in the West. In table 1.2, we compare the effects from own income and partner s income for male and female. We see that, The F-test result (p=0.0658) shows that for women the differenced income measures have no effect on their differenced satisfaction levels. And male care more about their labor income and depends less on their last period s satisfaction level. 17
And women only check their last period s satisfaction level in order to measure their current satisfaction for life. In the first specification, we see that the partner s income level in not as important as an individual s own income and have opposite signs. We may not treat this result as a verification of the collective model. By adding the last period s satisfaction level into the model, we get insignificant coefficients for the East. Maybe in the East, income is not the most important issue for measuring the SWB because of the economy type before the Unification. In the second specification, we add the reference incomes into the welfare function and by checking different combinations of reference group and income measures, we explore the sensitivity of the comparison income effect depends on the definition of the reference group. In Table 2.1, we see that for the full sample, the impact from the family reference income is higher than that from the individual reference income. This means individual s SWB is measured by the general performance of the family income. When we consider the number of kids as one of the reference measures, we see insignificant impact from the individual labor income. This shows that people will only compare their family income with the others when they have kids in the family. And the rank effect is also important to individual s SWB in a measure of family income. By looking at the log pseudolikelihood 18
levels, we see that the second definition of the reference group is the best among the three we mention here. But still we need to try more possible definitions in order to get a general view. When we compare the East and West differences in Table 2.2, we see similar results as those in Table 1 from the absolute income levels. For the comparison effects, we find that the West and the East residents only care about the family income rank. So we know that the ranking measure is more significant than the absolute reference income measures. The F-test result (p=0.2818) shows that in East the differenced absolute income measures have no effect on the differenced satisfaction levels. In this specification, we add several new covariates and the effects from individual labor income do not change very much. Only for the total sample, we see effects from the reference income. And we see effects from the ranking measurement for all of the three samples. So adding the ranking covariates, we may get a better measurement of the comparison income effect. Furthermore, in this specification we get the opposite results on the significance of the household comparison income effect compared with the ones in Ferrer-i-Carbonell (2005). In Ferrer-i-Carbonell (2005), they show that family reference income is only insignificant for the East. But after we introduce the labor reference income, we find that the absolute reference incomes are only significant for the total sample. We think the 19
reason of this could be that we add more income and comparison measures in the new welfare function and then separate the true effect from the household income effect. And also we get positive effect from the family reference income and this shows that their may be a tunnel effect on this comparison income effect. When people observe that other families have higher income level, they may expect themselves to be richer in the following period. In West sample, we get negative effect from partner s income, this can be taken as a support of the collective model. In the third specification, we introduce a new covariate but keep the basic analysis procedures used in the second specification. Here we use the distance between the individual income and the reference income. By using this new covariate, we get a better way to explain the comparison income effect, since the distance here is the main reason that people connected their SWB with the other individuals income levels. In Table 3.1, after we measure the reference income in a relative way through the distance, we get opposite directed impact for the family income reference except for the second type of the reference group. Here we also take this as a proof for the importance of including the labor income reference measures in the model. By looking at the log pseudolikelihood levels, we see that the third definition for the reference group is the best among the three we mention here. And the results are kind of robust among these reference groups. But still we need to try more possible definitions in order to get a general view. 20
Also, in this table we see different results about the comparison income effect comparing with the ones in Ferrer-i-Carbonell (2005). In Ferrer-i-Carbonell (2005), she gets significant impact from the distance covariate for only the full sample. We get only the ranking matters when an individual try to measure his SWB. The F-test result (p=0.5128) shows that in East the differenced income measures have no effect on the differenced satisfaction levels except the cumulative family inc. distribution. In the third specification, there is no big change on the effects from the labor income measures. And we only see significant effects from the family income ranking measurement in the total sample and the West sample. Since the log pseudolikelihood levels are not better than the ones from the second specification, we do not prefer this specification over the second one. In the fourth specification, instead of using one distance covariate, we use two of them which are covariates measure the absolute value for how much an individual is richer or poorer than the average person in the same reference group. The richer covariate measures the absolute value of how much the individual income is higher than the reference income. Similarly, the poorer covariate measures the absolute value of how much the individual income is lower than the reference income. Through these two new distances we can check whether the comparison income effects are symmetric or not. And further we can check whether for different income intervals, we have similar impact from the absolute income measures and the reference income measures. This would help 21
us to give some policy suggestion on taxation and social transfer. In Table 4.1 we find similar results as those in the third specification. The family income s rank is significant in all of the cases. So the introducing of ranking is important to the study of comparison income effect. By looking at the log pseudolikelihood levels, we see that the third definition for the reference group is the best among the three we mention here. And the results are kind of robust among these reference groups. But still we need to try more possible definitions in order to get a general view. In Table 4.2, the F-test result (p=0.5128) shows that in East the differenced income measures have no effect on the differenced satisfaction levels except the cumulative family inc. distribution. In this specification, there is no big change on the effects from the labor income measures. And we see significant effects from the family income ranking level in all the samples. Since the log pseudolikelihood levels are very closed to the ones from the second specification, we prefer this specification over the second one for its meaningful policy suggestions. When we compare the result here with the ones in Ferrer-i-Carbonell (2005), she finds out that being rich has no significant impact on SWB. Again we check the result from the individual labor income and the partner s income and get similar results as we have from the third specification. Since the family income ranking covariate has significant effect, we think the poorer and richer indicators may not be the main source of the change of SWB. 22
5.2 Other approaches to check the relative income effect 5.2.1 Using expected income to check the relative income effect After we discuss the results in Ferrer-i-Carbonell (2005) by doing some revisions, we try another approach to check the reference income effect which is used in Senik (2003). The estimation procedure named as the two-stage estimation is firstly estimating the predicted income of each individual by using a simple regression model with fixed effect. By using this estimation procedure, we do not need to take the trouble of choosing the best reference group. Then in the second stage, we substitute the predicted income levels into the welfare function to get the comparison income effect. Since the predicted incomes are highly correlated to the observed income levels, we need to use the bootstrap method in order to find out the reliable standard errors. In Table 5 we use two groups of covariates with and without the subjective measures to estimate the comparison income effect. From the results we see that when we only use a group of objective covariates to predict an individuals reference income, we get insignificant impact on SWB. But when we add a group of subjective measures in the first stage regression, we get significant effects from the reference incomes. We believe that the subjective measures capture some individual heterogeneity and help us see the size of comparison income effect more efficiently. By using this new approach, we find that only the last period s satisfaction level matters when we only use the objective measures to predict the reference income. But the reference income matters when we include the subjective measures in the first stage. 23
Also, we do not get significant effect from the labor income. It shows that as long as we construct a good reference group, the income effect can be measured only through the relative channel and the last period s satisfaction level may not have significant effect on SWB. The reason of that is after we include the subjective measures into the first step, we transfer the effect from the last period s SWB into the effect from the predicted income levels. The objective covariates used in the first stage are gender, age, marital status, number of children in household, number of adults in household, status of employment, years of education, partner s income, and region. And the subjective covariates used here are satisfaction levels with health, income, dwelling, job, leisure, housework, worried about economy development, savings, environment, peace, job security and child care. Furthermore we can use the decomposition methodology to check the share of contribution from objective and subjective measurements. 5.2.2 Parameter Stability Test for East and West Germany Table 6 examines the convergence process of comparison income effect in the East to the pattern in the West. Through comparing the annual East and West comparison effects, we see that the convergence in the last decade between the East and West is not that apparent. After reunification, for the West residents, being rich on the labor income level 24
mattes more than being poor. For the East residents, it seems that people are adopting the comparison income effect in the West and care more of being richer than the average labor income level. It may be caused by the economy transition and development in the East. This fact may be considered as a convergence of the comparison income effect. There is a similarity among these years income effects in the East and the West Germany. It seems that the East residences are following the patterns in the West and if we have the data available for the years after 2002, we may check this with more facts then. And we may find some convergence from the West to the East, too. The rank effect is not significant. This shows that in the same year, the position in the reference group may not have large impact on individual s SWB. 6. Conclusions This paper starts from the four specifications in Ferrer-i-Carbonell (2005), and take some further revisions to check the robustness of the reference group chosen in that paper. We find out that people not only measure their SWB through the absolute value of their individual labor income without caring much of their partner s labor income level, but also through the income levels of other individuals in the same reference group by the means of their relative positions in the group. 25
The further concern of this paper is that whether we can get a conclusion on which measure of income we should keep and which reference group we should refer. Also, we need to verify whether it is meaningful to keep some sub-satisfaction measures in the regression instead of the absolute income measures. Furthermore, if we believe that the reference group can not capture the gender and the region effects, we could try separate regressions by gender and region. 26
Reference: GA Akerlof and RE Kranton ECONOMICS AND IDENTITY, The Quarterly Journal of Economics, 2000. AV Banerjee A Simple Model of Herd Behavior, 1992. The Quarterly Journal of Economics, AE Clark Inequality-Aversion and Income Mobility: A Direct Test, Paris: DELTA. Mimeo, 2003. FA Cowell Measurement of Inequality 1998. E. Diener, EM. Suh and RE. Lucas. 1984. Subjective Well-being. HL Smith- Psychological Bulletin. E DIENER, EM SUH, RE LUCAS, and HL SMITH. Subjective well-being: Three decades of progress. Psychological bulletin, 1999. RA. Easterlin. 1995. Will Raising the Incomes of All Increase the Happiness of All? Journal of Economic Behavior and Organization. 27
RA. Easterlin. 2001. Income and Happiness: Towards a Unified Theory. The Economic Journal. A. Ferrer-i-Carbonell. 2002. Subjective questions to measure welfare and well-being. Tinbergen Institute Discussion Paper. A. Ferrer-i-Carbonell. 2005. Income and Well-being: An Empirical Analysis of the Comparison Income Effect. Journal of Public Economics, 89(5-6): 997-1019. BS Frey, A Stutzer What Can Economists Learn from Happiness Research? Journal of Economic Literature, 2002. P Frijters, JP Haisken-DeNew, MA Shields The Value of Reunification in Germany: An Analysis of Changes in Life Satisfaction, 2002. Paul Frijters, Ingo Geishecker, John P. Haisken-DeNew and Michael A. Shields INCOME AND LIFE SATISFACTION IN POST-TRANSITION RUSSIA: A NEW EMPIRICAL METHODOLOGY FOR PANEL DATA. B. Headey, R. Veenhoven and A Wearing. 1991. Top-down Versus Bottom-up Theories of Subjective Well-being. Social Indicators Research. 28
A Kapteyn, S van de Geer, H Van de Stadt Interdependent preferences: an econometric analysis, Journal of Applied Econometrics, 1997. Lundberg, S. J., Pollak, R.A. and Wales, T.J. Do husbands and wives pool their resources? Evidence from the United Kingdom child benefit. The Journal of Human Resources v. 32 (Summer 1997) p. 463-80. M McBride Relative-income effects on subjective well-being in the cross-section, Journal of Economic Behavior and Organization, 2001. Yair Mundlak On the Pooling of Time Series and Cross Section Data, Econometrica, Vol. 46, No. 1 (Jan., 1978), pp. 69-85. AJ Oswald, AE Clarke Satisfaction and Comparison Income, Journal of Public Economics, 1996. B.M.S. van Praag, P. Frijters and A. Ferrer-i-Carbonell. 2000. A Structural Model of Well-being : with an application to German Data. Tinbergen Institute Discussion Paper. CLAUDIA SENIK. When Information Dominates Comparison. Learning from Russian Subjective Panel Data. William Davidson Institute Working Paper No. 495, 2003. 29
C Senik What Can We Learn from Subjective Data: the Case of Income and Well- Being, Journal of Economic Surveys, 2005. Huib van de Stadt, Arie Kapteyn, Sara van de Geer The Relativity of Utility: Evidence from Panel Data, Review of Economics and Statistics, Vol. 67, No. 2 (May, 1985), pp. 179-187. 30
Table 1.1 IV for lagged differenced general satisfaction and individual labor income, General Satisfaction, first specification First Differenced Ordered Logit, GSOEP 1991 2002 Total Westerners Easterners Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio ln(labor income) -.0300323-2.71 -.0292483-2.14 -.0295422-1.56 ln(partner s income).004771 1.09.0011749 0.21.0105921 1.49 ln(last period s SWB).3182264 3.79.4414267 3.99.123624 0.96 Number observations 34871 23546 11325 Number of individuals 7825 5450 2400 Log pseudolikelihood -61549.28-41294.67-20204.043 31
Table 1.2 IV for lagged differenced general satisfaction and the individual labor income, General Satisfaction, first specification First Differenced Ordered Logit, GSOEP 1991 2002 Total Male Female Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio ln(labor income) -.0300323-2.71 -.0404495-2.23 -.0240265-1.72 ln(partner s income).004771 1.09.0016584 0.28.0107045 1.60 ln(last period s SWB).3182264 3.79.2670388 2.14.3599035 3.19 Number observations 34871 17452 17419 Number of individuals 7825 3921 3904 Log pseudolikelihood -61549.28-30589.461-30929.829 32
Comparing with the results in Ferrer-i-Carbonell (2005) 33
Table 2.1 IV for lagged differenced general satisfaction and the individual labor income, General Satisfaction, second specification First Differenced Ordered Logit, GSOEP 1991 2002 First Second Third Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio ln(family reference income).4176637 2.30.4887631 3.06.4256025 3.34 ln(labor reference income) -.028929-2.03 -.0427861-3.84 -.0097856-0.85 ln(cumulative family inc. distribution).1029192 4.90.0877754 4.40.1022941 4.91 ln(cumulative labor inc. distribution).0059394 0.43 -.0091211-0.75.0174236 1.25 Log pseudolikelihood -61526.914-61523.561-61525.304 Note: The First group of reference characteristics is age, years of education, gender and region. The Second group of reference characteristics is age, years of education, gender, employment status and region. The Third group of reference characteristics is age, years of education, gender, the number of children in the family and region. 34
Table 2.2 IV for lagged differenced general satisfaction and the individual labor income, General Satisfaction, second specification First Differenced Ordered Logit, GSOEP 1991 2002 Total Westerners Easterners Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio ln(labor income) -.029313-2.65 -.0281744-2.07 -.0293717-1.55 ln(partner s income).0003047 0.07 -.0034412-0.61.0062597 0.87 ln(family reference income).4176637 2.30.4184114 1.88.3150232 0.96 ln(labor reference income) -.028929-2.03 -.029454-1.21 -.0245758-1.38 ln(cumulative family inc. distribution).1029192 4.90.0967175 3.77.1214547 3.32 ln(cumulative labor inc. distribution).0059394 0.43.0076997 0.46.0047956 0.19 ln(last period s SWB).3142796 3.74.4379935 3.96.1173341 0.91 Number observations 34871 23546 11325 Number of individuals 7825 5450 2400 Log pseudolikelihood -61526.914-41282.234-20193.202 35
Comparing with the results in Ferrer-i-Carbonell (2005) 36
Table 3.1 IV for lagged differenced general satisfaction and the individual labor income, General Satisfaction, third specification First Differenced Ordered Logit, GSOEP 1991 2002 First Second Third Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio ln(fam. inc.)- ln(fam. ref. inc.) -.0175706-0.23.0450718 0.61 -.0519654-0.75 ln(labor inc.)- ln(labor ref. inc.) -.0040037-0.81.004194 0.89 -.0064695-1.32 ln(cumulative family inc. distribution).1046705 3.32.0643143 2.27.1094063 3.73 ln(cumulative labor inc. distribution).010009 0.66 -.0115159-0.90.0228647 1.51 Log pseudolikelihood -61530.651-61536.442-61530.649 Note: The First group of reference characteristics is age, years of education, gender and region. The Second group of reference characteristics is age, years of education, gender, employment status and region. The Third group of reference characteristics is age, years of education, gender, the number of children in the family and region. 37
Table 3.2 IV for lagged differenced general satisfaction and the individual labor income, General Satisfaction, third specification First Differenced Ordered Logit, GSOEP 1991 2002 Total Westerners Easterners Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio ln(labor income) -.0304787-2.80 -.0297766-2.22 -.0294754-1.57 ln(partner s income).0005917 0.13 -.003024-0.54.0060584 0.84 ln(fam. inc.)- ln(fam. ref. inc.) -.0175706-0.23 -.0530179-0.63.1773781 1.01 ln(labor inc.)- ln(labor ref. inc.) -.0040037-0.81 -.0052569-0.84 -.0006984-0.09 ln(cumulative family inc. distribution).1046705 3.32.1111409 2.94.0713557 1.23 ln(cumulative labor inc. distribution).010009 0.66.0123992 0.69.0053986 0.19 ln(last period s SWB).3131531 3.73.4386356 3.96.114851 0.89 Number observations 34871 23546 11325 Number of individuals 7825 5450 2400 Log pseudolikelihood -61530.651-41283.614-20193.911 38
Comparing with the results in Ferrer-i-Carbonell (2005) 39
Table 4.1 IV for lagged differenced general satisfaction and the individual labor income, General Satisfaction, fourth specification First Differenced Ordered Logit, GSOEP 1991 2002 First Second Third Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio richer in fam. inc. -.0121196-0.13.0771754 0.86 -.0590383-0.66 richer in labor inc. -.0425308-1.12 -.0303888-0.97 -.0463005-1.28 poorer in fam. inc. -.0073019-0.06 -.0316201-0.30.0176391 0.18 poorer in labor inc..0029017 0.57 -.0077621-1.34.00516 1.01 ln(cumulative fam. inc. distribution).1003352 2.71.067791 2.11.1038324 3.17 ln(cumulative labor inc. distribution).0108298 0.72 -.0123777-0.96.0239357 1.58 Log pseudolikelihood -61530.045-61535.551-61529.809 Note: The First group of reference characteristics is age, years of education, gender and region. The Second group of reference characteristics is age, years of education, gender, employment status and region. The Third group of reference characteristics is age, years of education, gender, the number of children in the family and region. 40
Table 4.2 IV for lagged differenced general satisfaction and the individual labor income, General Satisfaction, fourth specification First Differenced Ordered Logit, GSOEP 1991 2002 Total Westerners Easterners Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio ln(labor income) -.0302319-2.77 -.0295292-2.20 -.0292263-1.56 ln(partner s income).0005208 0.12 -.0027076-0.48.0059444 0.83 richer in fam. inc. -.0121196-0.13.0771754 0.86 -.0590383-0.66 richer in labor inc. -.0425308-1.12 -.0303888-0.97 -.0463005-1.28 poorer in fam. inc. -.0073019-0.06 -.0316201-0.30.0176391 0.18 poorer in labor inc..0029017 0.57 -.0077621-1.34.00516 1.01 ln(cumulative fam. inc. distribution).1003352 2.71.067791 2.11.1038324 3.17 ln(cumulative labor inc. distribution).0108298 0.72 -.0123777-0.96.0239357 1.58 ln(last period s SWB).3135002 3.73.439795 3.97.1168543 0.91 Number observations 34871 23546 11325 Number of individuals 7825 5450 2400 Log pseudolikelihood -61530.045-41282.802-20193.028 41
Comparing with the results in Ferrer-i-Carbonell (2005) 42
Table 5 IV for lagged differenced general satisfaction and the individual labor income, General Satisfaction, Two-stage estimation for robustness First Differenced Ordered Logit, GSOEP 1991 2002 Objective covariates only Objective and subjective covariates Coefficient t-ratio Coefficient t-ratio ln(labor income) -.0111586-0.42.0123201 0.36 ln(partner s income).0083883 1.31.0162015 1.83 predicted family income -.2278976-0.24 11.47973 12.00 predicted labor income.4388762 1.50 -.9309873-2.89 ln(last period s SWB).296333 2.27.3197933 1.79 Number observations 17700 8935 Number of individuals 4757 3097 Log pseudolikelihood -30002.913-15168.215 43