Social Interactions in Job Satisfaction

Transcription

1 Social Interactions in Job Satisfaction Semih Tumen Central Bank of the Republic of Turkey Tugba Zeydanli Paris School of Economics October 15, 2012 Abstract We test the existence of spillover externalities in job satisfaction in the sense that the individual workers job satisfaction is increasing in the overall job satisfaction in the reference group. If there exist positive spillover externalities, then there are increasing returns to social and economic policies to raise job satisfaction. We use the British Household Panel Survey (2008, the latest release of BHPS data) to investigate whether social interactions play a role in the determination of individual-level job satisfaction. Our econometric framework distinguishes contextual effects from endogenous effects within a canonical linear-in-means model of social interactions; that is, our empirical model recognizes the reflection problem. Using an instrumental variables strategy motivated by the reflection problem, we conclude that there exist significant endogenous social interactions in job satisfaction in the UK. This is a manifestation of complementarities between individuallevel and group-level job satisfaction levels. We also report a rich set of novel estimates on the effects of group-level contextual variables such as the group means of education, demographic characteristics, worker preferences, and job- and employer-related characteristics on individual-level job satisfaction. JEL codes: C31, J28. Keywords: Job satisfaction; social interactions; reflection problem; British Household Panel Survey. We thank Andrew Clark and Claudia Senik for useful comments and suggestions. The views expressed here are of our own and do not necessarily reflect those of the Central Bank of the Republic of Turkey. All errors are ours. semih.tumen@tcmb.gov.tr. Central Bank of the Republic of Turkey, Istiklal Cad. No:10, Ulus, Ankara, Turkey. Tugba.Zeydanli@malix.univ-paris1.fr. CES-Centre d Economie de la Sorbonne Maison des Sciences Eco boulevard de l Hôpital Paris cedex 13.

2 1 Introduction Job satisfaction conveys useful information about economic life that should not be ignored [Freeman (1978)]. It has extensive behavioral consequences. For example, job satisfaction is a significant determinant of labor market mobility and, in particular, the quitting behavior [e.g., Akerlof, Rose, and Yellen (1988) and Clark, Georgellis, and Sanfey (1998)]. It is also documented to have a strong relationship with the relative income structure among peer groups [e.g., Clark and Oswald (1996) and Card, Mas, Moretti, and Saez (2010)]. Moreover, it is shown to have a positive correlation with worker performance and labor productivity [e.g., Argyle (2001) and Oswald, Proto, and Sgroi (2012)]. Although several aspects of job satisfaction have been extensively studied in the empirical literature, whether there exist spillover externalities in job satisfaction in the sense that the individual workers job satisfaction is increasing in the overall job satisfaction in the reference group remains as an unanswered question. If indeed there exist positive spillover externalities, then there are increasing returns to social and economic policies to raise the job satisfaction level within the worker population. Our ultimate goal in this paper is to investigate if there exist any visible footprints of social interactions in job satisfaction. This will be the first paper attempting to answer this question in the literature. Our empirical analysis of social interactions in job satisfaction is within the context of spillovers and associated economic models. Spillovers are classic examples of nonpecuniary externalities that affect individual-level outcomes [Arrow and Hahn (1971) and Becker and Murphy (2000)]. There is a vast literature studying positive spillovers in the sense that social interactions induce a similarity in outcomes across the members of a reference group. 1 Our econometric model will be based on a canonical linear-in-means framework. 2 This model is a base for the bulk of empirical work on social interactions. There is a common mistake often made in this literature: the contextual effects are typically not clearly distinguished from the endogenous social 1 Cooper and John (1988) and Milgrom and Roberts (1990) are excellent introductory readings for the literature on social interactions and the associated spillover externalities. 2 See Blume, Brock, Durlauf, and Ioannides (2011) for an in-depth background information on linear-in-means models, including a comprehensive discussion on microfoundations and econometric identification. Also see Blume and Durlauf (2001), Brock and Durlauf (2001b), and Soetevent (2006). 2

3 effects. In other words, the literature often fails to distinguish the effect that comes from predetermined group-level differences from the effect coming from endogenous (e.g., behavioral) differences. This confusion leads to the well-known reflection problem [Manski (1993, 2000)]. Following Manski s prescription of econometric identification, we employ an appropriate strategy to estimate social interactions in job satisfaction. In particular, we distinguish the effect coming from exogenous group-level characteristics (such as group-level means of demographic and job/employer related variables) from the effect coming from group-level outcome variable (i.e., the mean job satisfaction in the reference group). As an example, suppose for a moment that the individual-level job satisfaction is higher both because the group-level job satisfaction and, say, the group-level promotion opportunities are higher. Then the effect coming from the group-level job satisfaction cannot be distinguished from that coming from group-level promotion opportunities, unless an appropriately designed identification strategy is used. We employ such a strategy to distinguish between these two forces. At the center of our identification strategy lies an instrumental variables argument motivated by an exclusion restriction. More specifically, we show that there must be at least one individual-level characteristic, which does not correspond to a contextual effect when averaged out. 3 The British Household Panel Survey (BHPS 2008 release), the dataset that we use in our empirical analysis, offers such a candidate for such an instrument: the day of the week. 4 There is an emerging literature investigating if there is any correlation between the day of interview and the level of job satisfaction self-reported by the respondents. 5 We exploit these correlations to construct a variable that rationalizes the use of an instrumental variables strategy to resolve the reflection problem. 6 We use industry/region combinations as we describe in Section 2.2 to construct the underlying reference groups that the interactions can potentially be effective. Our main findings can be classified into two categories: the results pertaining to endogenous 3 See Section 3 for a detailed explanation. 4 The BHPS records the date of each interview as day-month-year, allowing to compute the day-of-the-week on which the interview occurs. 5 See Taylor (2006), Akay and Martinsson (2009), and Helliwell and Wang (2011) for recent studies. 6 Section 3.1 presents a comprehensive theoretical discussion on the role of instrumenting in linear-in-means model. Section 3.2 how we accordingly construct our instrument using the day-of-the-week idea. 3

4 effects and those pertaining to contextual effects. We document that there are significant endogenous social effects in job satisfaction in the UK. 7 The crux of the matter is the following: each worker s job satisfaction level is determined not only by exogenous individual- and grouplevel characteristics, but also by the job satisfaction levels of the other people around. This suggests significant positive spillover externalities in job satisfaction; that is, group-level and individual-level job satisfaction are complementary. Our paper is the first in this literature to report endogenous social effects in job satisfaction, thus there are no comparable papers in terms of the results regarding the endogenous social effects. Our methodology of peer group construction, however, can be compared to several related but distinct strands in the social interactions literature. We group people in terms of the industry-region combinations reported in the BHPS. This is similar to the papers including Benabou (1993), Gabszewicz and Thisse (1996), and Glaeser, Sacerdote, and Scheinkman (1996), who also form their peer groups in terms of industry-region combinations or only in terms of regional proximity. We abstract from other peer group design approaches such as group formation based on random assignment [see, e.g., Ioannides (1990)], individual incentives [see, e.g., Bala and Goyal (2000)], coalition games [see, e.g., Myerson (1991)], and Tiebout s equilibrium [see, e.g., Bewley (1981)]. The result that group-level job satisfaction positively affects individual-level job satisfaction has important policy implications. This result means that there are increasing returns to social and economic policies to raise job satisfaction. Our results for the existence of contextual social effects in job satisfaction can be further assessed in four sub-categories. First, we find that an individual s job satisfaction is higher if she is above the median earnings level in her reference (or peer) group. This finding is consistent with the findings reported in the relative income comparisons literature [see, e.g., Clark and Oswald (1996), Hamermesh (2001), and Card, Mas, Moretti, and Saez (2010)]. Our paper differs from the others in the sense that we extract our results from a canonical semistructural empirical model of social interactions. In addition, we present a new result: we show that whether the group-level (average) earnings are below or above the median earnings 7 The result that there are significant endogenous social effects in job satisfaction is related to a large body of research in the social interactions literature. See Blume, Brock, Durlauf, and Ioannides (2011) for an up to date review of the most recent literature. 4

5 level has no statistically significant effect on individual-level job satisfaction. In other words, the worker cares her own position on the earnings distribution rather than the position of the mean earnings of the group she is associated with. It is possible to associate these findings with the well-known Easterlin paradox [Easterlin (1974)]. According to Easterlin, individuals with higher income levels tend to report higher subjective well-being in a given country. At the international level, however, self-reported subjective well-being does not vary significantly with country-level per capita GDP. Our findings can be regarded as the group-level analogue of Easterlin s results. Second, we find that the coefficients measuring the effects of individual-level observed characteristics on job satisfaction and the coefficients of their group-level counterparts can alternate signs. This is a rather surprising finding. For example, it is well-known in the empirical literature that female workers are more satisfied jobwise. Our estimations confirm this result. But we also find that individual-level job satisfaction is higher in groups with a greater fraction of male workers. More educated workers are less satisfied with their jobs. But proximity to a group of more educated workers raises individual-level job satisfaction. Similarly, being married is positively correlated with job satisfaction, whereas working in a group with a greater fraction of married worker decreases individual-level job satisfaction. Workers with satisfactory health conditions are less satisfied with their jobs, while working in a group with a larger fraction of workers reporting satisfactory health improves individual-level job satisfaction. Finally, preference to work fewer hours is negatively correlated with individual-level job satisfaction, but working close to a larger fraction of peers with preference to work fewer hours is positively correlated with individual-level job satisfaction. Third, there is a set of variables for which the individual-level and group-level counterparts produce coefficients of the same sign. For example, older workers are less satisfied jobwise and working a group of older workers decreases individual-level job satisfaction. Working under a temporary contract produces a negative coefficient both at the individual and group levels. Similarly, being a public-sector worker and working close to a greater fraction of public-sector workers is negatively associated with individual-level job satisfaction. Finally, having access 5

6 to promotion opportunities increases job satisfaction both at the individual and group levels. Fourth, there is another set of variables for which the individual-level coefficients are statistically significant, while the group-level counterparts turn insignificant. These variables are listed as follows: being single, working in a small-size firm, being a union member, preference to work more hours, having very good health conditions, and, finally (as we mention above), getting paid more than the median income level in the reference group. The plan of the paper is as follows. Section 2 provides an overview of our data along with summary statistics. Section 3 explains the details of the econometric model and the instrumental variables strategy we use. Section 4 presents the estimates and discusses the results summarized above in detail. Section 5 concludes. 2 Data This section summarizes the data that we work with (Section 2.1), performs a preliminary regression analysis using a random-effects ordered probit model to understand the basic correlations across the variables that we use in the dataset (Section 2.2), and describes the construction of the reference (or peer) group variables that we use in our empirical investigation of social interactions in job satisfaction (Section 2.3). 2.1 Basic Facts The main source of the data is the British Household Panel Survey (BHPS). The BHPS provides information on individual, household, and job-related characteristics starting from 1991 to 2008 in the Great Britain, Scotland, Wales, and Northern Ireland. It yearly follows the same representative sample of households interviewing every adult member of sampled households and assigns a unique identification number for each interviewer. The date of each interview is recorded as day-month-year; the day-of-the-week on which an interview occurs. We use an unbalanced panel data, which the individuals can enter and leave the estimating sample over time. Eighteen waves of data are available. Due to changes in the measurement 6

7 instrument, in wave-1, the job satisfaction scores are higher than those in other waves [Rose (1999)]. Therefore, we accordingly drop wave-1 from our analysis and we use the data from wave-2 to wave-18. The individual-level job satisfaction in the BHPS dataset is reported based on a seven-point scale ranging from 1 (not satisfied at all) to 7 (completely satisfied). At each and every interview, the employed workers are asked to rate the job satisfaction level regarding of promotion prospects, total income, relationship with boss, job security, able to use their initiatives in the work, the actual work itself, and hours worked. The last question about job satisfaction is Overall, how satisfied or dissatisfied are you with your present job?, which is again measured on the 1 7 scale and named the overall job satisfaction. This is a direct measure of individuals utility derived from their current job [Clark and Oswald (1996)]. The dependent variable is the overall job satisfaction in our analysis. For the individual- and job-related characteristics, we follow the recent job satisfaction studies using the BHPS and control for gender, age, education levels, preferences over working hours, types of contract, size of establishment, promotion opportunities, union membership, and health status [Taylor (2006)]. We collapse the education-levels into seven broad groups as follows: higher degree refers to postgraduate education, first degree refers to college education, A -level, O -level and other higher qualification refer to high school graduates of different types (consistent with the UK education system), vocational qualification refers to teaching, nursing, commercial, apprenticeship, and the certificate of secondary education (CSE), and, finally, the ones with no qualification. We also construct a dummy variable ( Income ) for earnings. It is equal to 1 if the worker earns more than the median level of earnings in her reference group and within the corresponding wave and it is equal to 0 otherwise. How we construct the reference groups are described in detail in the next subsection. Table (1) presents the summary statistics of the data that we use in our analysis. In order to be in our sample, the respondent must report the overall job satisfaction rate. 8 The mean age of the respondents is about 39. Among the 98,148 observations, 49% are male, 56% are married, 33% are single, 3.1% have higher degree, 13.3% 8 We also drop the ones reported 0 job satisfaction, as the scale of job satisfaction is 1 7 in the BHPS. 7

8 have first degree, another 13% have A -level degree, 20.1% have O -level degree, 28.4% have other higher qualifications, 10.9% have vocational qualifications, and the remaining 11.2% have no qualifications. 2.4% and 1.3% have temporary and fixed term contract, respectively. 17.3% work in the public sector and 69.1% work in a company of size 200 workers or larger. 24% are union members. 7.9% prefer to work more hours and 31.3% prefer to work fewer hours. 25.1% report their health to be very good, whereas 16.4% report to be satisfactory. 50.7% earn above the median monthly income. The mean overall job satisfaction rate is 5.37 out of 7, with a standard deviation of Notice that there is another dummy variable in the table labeled Tuesday-Saturday-Sunday. It is equal to 1 if the interview is conducted on a Tuesday, Saturday, or Sunday and it is equal to 0 otherwise. This dummy variable will be served as an instrument in our empirical analysis of social interactions. 31.7% of the workers in our sample are interviewed on Tuesday, Saturday, or Sunday. See Sections 2.3 and 3 for the details of instrument construction. All the means and the standard deviations reported in Table (1) are calculated using the BHPS frequency weights Reference Groups Since our primary objective is to investigate endogenous and contextual social interactions effects in job satisfaction, we construct reference groups in which workers can potentially engage in interactions and in which social forces can operate effectively. Following the papers in the social interactions literature including Gabszewicz and Thisse (1996) and Glaeser, Sacerdote, and Scheinkman (1996) we construct our reference groups as industry and region combinations. We cluster the geographical locations in the following eleven regions: 1) London, 2) South East, 3) South West, 4) East Anglia, 5) East Midlands, 6) West Midlands, 7) North West, 8) North East, 9) Yorkshire & Humberside, 10) Wales, and 11) Scotland. 10 Nine industry categories are selected at the one-digit level as follows: 1) energy & water supplies, 2) extraction of minerals & manufacture of metal goods, mineral products & chemicals, 3) metal goods, engineering & vehicles, 4) other manufacturing industries, 5) construction, 6) 9 Until wave 11, for the original BHPS sample, the weight is wlrwght. Afterwards, for all samples, wlrwtuk1 is used. Our analysis also takes into consideration the complex survey design [see BHPS, Volume A, 2 5]. 10 We exclude the Northern Ireland due to clustering issues [see BHPS, Volume A, 2 5]. 8

9 distribution, hotels & catering (repairs), 7) transport & communication, 8) banking, finance, insurance, business services & leasing, and 9) other services. Thus, there are 99 reference groups in our sample. This structure assumes that, say, workers in the banking industry located in Scotland are exposed to different social interactions effects than, say, workers in the metal goods industry located in East Anglia. An alternative peer group selection would be to use occupation groups rather than industries. But, this would mean that a plumber in the banking industry would be exposed to the same social interactions effects as a plumber in the construction industry, which would perhaps not be the case. It is well-known that occupation-level groupings may create extra noise in large datasets. We follow the conventional wisdom and stick with the industry-region combinations in formulating our reference groups. As we discuss in Sections 3 and 4, we include the group-level weighted averages of the observed characteristics, summarized in Table (1), as control variables into our social interactions regressions. Therefore, we will have group-level variation as well as individual-level variation in our analysis. 2.3 The Day of the Week In this section, we perform a preliminary analysis of the main determinants of individual-level job satisfaction. The main purpose is to understand the basic correlations in our dataset and to have a suggestive opinion on the sources of individul-level job satisfaction. To our knowledge, this is the first paper to use the 2008 release of the BHPS data in analyzing job satisfaction. As a dependent variable, job satisfaction has an ordered nature, where it takes a value between 1 and 7. It is specified to be a function of the demographic and household characteristics as well as job- and employer-related characterictics. Time and industry dummies are included. We also include dummy variables for the day of the interview (i.e., the day of the week on which the interview is conducted). Following Clark and Oswald (1996), Clark (1997), and Taylor (2006), we run a random- 9

10 effects ordered probit regression. See Table (2) for a complete list of control variables and the regression results. Consistent with the findings in the previous literature, we find that job satisfaction is higher for females (than males), younger workers (than older ones), married (than nonmarried), and less educated (than more educated) workers controlling for time, industry, and job- and employer-related characteristics. We also report correlations regarding the day-of-the-week effect. Only the two of the day dummies produce statistically significant coefficients: Friday and Saturday. More precisely, we find that Saturday and Friday are the days on which the self-reported job satisfaction level is the highest. Our results suggests that Sunday/Monday/Wednesday/Thursday are the days on which the level of self-reported job satisfaction is the lowest, although these results are statistically insignificant. Combining our findings with the main results in the related literature, we conclude that Friday/Saturday are the days on which job satisfaction is highest and Sunday/Monday/Wednesday are the ones on which it is lowest. We will use this result in Section 3, when we construct our instrumental variable to estimate social interactions. 3 Econometric Framework The econometric framework that we use in this paper is a canonical linear-in-means model of social interactions. 11 Our ultimate goal is to estimate social interactions in job satisfaction. In particular, we would like to estimate the effects of (1) group-level job satisfaction the endogenous social effect and (2) group-level exogenous characteristics the contextual effect on individual-level job satisfaction. The linear-in-means model of social interactions is plagued with the well-known reflection problem [Manski (1993)]. Under the reflection problem, social interactions cannot be identified in a linear-in-means model. The simplest way to resolve this issue is to use an appropriately formulated instrumental variables strategy The standard empirical tool to analyze the determinants of job satisfaction is the ordered-choice model. Here we prefer a linear model, since a framework for identifying social interactions in an ordered-choice model does not exist in the literature, although there are clear-cut results to identify social interactions within binary [Brock and Durlauf (2001a)] or multinomial [Brock and Durlauf (2002)] discrete choice models. For theoretical attempts to model social interactions within an ordered-choice framework, see Aradillas-Lopez (2011) and Tumen (2011). 12 When an instrument is not available, it is necessary to invoke nonlinearities to identify social interactions. See Brock and Durlauf (2001a) for an example. 10

11 This section provides a detailed description of our econometric model for the purpose of familiarizing the reader with the basic concepts we frequently mention throughout the paper. Section 3.1 presents our empirical model and the associated technical issues (i.e., the reflection problem) including a formal statement about the need for a valid instrument to identify social interactions. Section 3.2 describes how we construct our instrument using the BHPS data, a detailed overview of which is given in Section The Empirical Model Each individual i I is a member of a group g G, where I is the number of individuals in the worker population and G is the number of groups, where I > G. The following linear-in-means model will be at the core of our empirical analysis: ω ig = β 0 + β 1 X ig + β 2 Y g + Jm g + ɛ ig, (3.1) where the dependent variable, ω ig, is the individual-level outcome of observation i in group g, X ig is a vector of individual-level observed characteristics of i in group g, Y g is a vector of group-level observed characteristics of group g, m g = E[ω ig g, F ig ] is the mean outcome in group g, and ɛ ig is a random error term with E[ɛ ig g, F ig ] = 0. In our notation, F ig corresponds to the distribution of individuals in group g and this distribution is possibly different for each group. The distinction between β 2 (contextual effects) and J (endogenous effect) is the key notion in this model. The former measures the effect of exogenous group-level variables on the individual-level outcome, while the latter measures the effect of endogenous group-level outcome on the individual-level outcome. Our ultimate goal is to clearly distinguish β 2 from J and to separately identify the effects of group-level variables on the individual-level outcome. However, econometric identification is a problematic issue in this standard setting. To define the identification problem, we take the conditional mathematical expectations of the both sides of the linear-in-means equation, where the conditioning is on g and F ig, for all 11

12 i and g. This gives us m g = β 0 + β 1 X g + β 2 Y g + Jm g, (3.2) where X g = E[X ig g, F ig ]. X g can be named as the group-level mean of individual-level observed characteristics and it may or may not coincide with Y g. Notice that m g appears in both sides of Equation (3.2). Solving for m g yields the result that m g = β 0 1 J + β 1 1 J X g + β 2 1 J Y g. (3.3) The reflection problem states that if the dimensions of the vectors X g and Y g are the same, then linearity masks the econometric identification of the (endogenous) social interactions parameter J. To formalize this statement, we plug Equation (3.3) into Equation (3.1), which gives us the estimating equation ω ig = β 0 1 J + β 1X ig + Jβ 1 1 J X g + β 2 1 J Y g + ɛ ig. (3.4) When the reflection problem is in effect, J and β 2 cannot be distinguished from each other, which implies that social interactions cannot be identified. To see this, set X g = Y g, which yields the equation ω ig = β 0 1 J + β 1X ig + Jβ 1 + β 2 1 J Y g + ɛ ig. (3.5) It is obvious that, in this equation, it is impossible to separate J from β 2 econometrically. One solution is the existence of an additional X g which is not in Y g. If such a X g exists, then social interactions (J) and also all the other model parameters are identified by applying simple ordinary least-squares method on Equation (3.4). In other words, one individual-level variable, the mean of which cannot be regarded as a group-level variable, is required for identification of social interactions. 12

13 To demonstrate this, let X g be an element of the set X g and let β 1 be the coefficient associated with X g. Let X g / Y g. Then, Equation 3.4 can be rewritten as ω ig = β 0 1 J + β 1X ig + J β 1 1 J X g + β 2 1 J Y g + ɛ ig. (3.6) Clearly, β1 is an element of the vector of paramaters β 1. From Equation (3.6), β1 can be identified, which implies that J and β 2 can separately be identified using this framework. The key point is the existence of a variable X g, which does not correspond to a contextual effect. Individual-level variables such as gender, education, age, marital status, and so on necessarily correspond to contextual effects when averaged out. One should find an individuallevel variable X g such that it cannot be interpreted as a group-level characteristic. If such a variable exists, it serves as an instrument and secures identification of J and β 2 separately. The BHPS data allows us to construct such an instrument. In the next subsection, we describe how we construct our instrument and implement our strategy. 3.2 Construction of the Instrument The date of each interview is recorded in the BHPS. In the analysis of subjective well-being, whether or not the day of the week on which the interview is conducted has a significant effect on self-reported job satisfaction levels has become an active line of research in recent years. In this literature, Friday/Saturday have been typically reported to be the days with the highest job satisfaction responses and Sunday/Wednesday/Monday to be the lowest. In the rest of this subsection, we describe how we use the day-of-the-week idea to construct our instrument, X g. We construct a binary variable D by assigning D = 1 for those who are interviewed on Tuesday, Saturday, or Sunday and D = 0 otherwise. 13 Three fundamental factors drive this choice of the days in constructing D: 1. D has to predict the job satisfaction score reasonably well. 13 See below for additional explanation on how we choose these days and for robustness analysis under alternative day selections. 13

14 2. The day groups has to be constructed in such a way that it has to be free of the selfselection problem. For example, workers interviewed on Friday or Saturday may be more likely to be satisfied with their job since they like to work hard during the weekdays and weekends are the only available time for them to file their response to the survey. 14 A classical endogeneity problem may arise if this is not paid attention. Our construction takes this consideration into account. 3. It has to justify the exlusion restriction; that is, it should not reflect a contextual effect when it is averaged out in the reference group. Clearly, the mean of D in the reference group does not correspond to a context, whereas variables like age, education, gender, and so on do represent a context. If these three criteria are met, then D is a reasonable candidate to be used in the construction of our instrument. The formal definition of Xg is as follows: X g = E [ D ig g, F ig ]. (3.7) In other words, Xg is the group-level counterpart (i.e., the mean) of the individual-level day dummy D i in each group. Clearly, Xg [0, 1], g G. The choice of the days is the crucial issue in this construction. As we mention above, D = 1 if the interview is conducted on Tuesday, Saturday, and Sunday and D = 0 otherwise. The regression results we present in Section 4 confirm that the first criterion is met. To satisfy the second criterion, it is necessary to construct D in such a way that the selectivity issue is avoided. The results reported in the literature as well as our estimations from the randomeffects ordered probit model we document in Section 2 provide us guidance on this point. Friday and Saturday are the best days and Sunday/Monday/Wednesday are the worst ones in terms of self-reported job satisfaction. The dummy variable has to constructed in such a fashion that the D = 1 category should not oversample the good days or the bad days. For example, putting workers surveyed in Fridays and Saturdays together into the D = 1 may lead 14 Most studies in the literature do not pay attention to this factor. See our companion paper, Tumen and Zeydanli (2012), for day-of-the-week effect estimates accounting for selectivity. 14

15 to a selection problem as described above, in criterion 2. For this reason, we pick one from the good days, one from the bad days, and one from the intermediate days to construct D. We try a large number of alternative configurations [see Table (5)] and decide that the best combination the criteria is: D = 1 if the survey is taken on Tuesday, Saturday, or Sunday; and D = 0 otherwise. The third criterion holds by definition. The group mean of D, which is X g, cannot be regarded as a group-level characteristic, which means that its coefficient does not correspond to a contextual effect. We perform estimations over this setup and the results are reported in the next section. 4 Results and Discussion We begin in Table (3) by presenting the estimates for the regression model given in Equation (3.6). The coefficients reported in the table are first-stage results and estimates for the grouplevel characteristics, given in the bottom half of the table, should not directly be interpreted as social interactions. Using the identifying formulas given in Section 3, we derive a set of translated (or implied) coefficients that indeed correspond to social interactions estimates. These translated effects are presented in Table (4). Parallel to the decomposition idea that our identification strategy rests on, the results of these two tables can be evaluated in two broad categories: endogenous social effects and contextual effects. In the rest of this section, we provide a variable-by-variable assessment of our findings. 4.1 Estimates for Endogenous Social Interactions A group-level variable is endogenous if its individual-level counterpart is a choice variable. Hence, the associated group-level variable can be defined as the effect of other people s behavior of individual-level behavior. This a classic example of spillover externalities. Our findings see the first line in Table (4) verify that there significant positive spillover externalities in job satisfaction; that is, the group-level (i.e., mean) job satisfaction is positively related to individual-level job satisfaction. To put it differently, an individual worker s job satisfaction level tend to be higher in a group of workers who are highly satisfied jobwise. 15

16 The reported coefficient is around 1.05, which means that a one-unit increase in group-level job satisfaction leads to a more than one-unit increase in individual-level job satisfaction, on average. This result suggests that there are huge gains to policy interventions to increase individual-level job satisfaction as there are large positive feedback effects from group-level job satisfaction toward individual-level job satisfaction in the form of spillover externalities. In particular, it is reported in the literature that job satisfaction is positively related to worker productivity. This means that existence of spillover externalities introduces increasing returns (in production) to raising job satisfaction at the individual level. It has also implications for discussions of optimal design of worker groups and teamwork as bringing together highly satisfied workers in production processes may lead to more efficient outcomes in the existence of complementarities between group-level and individual-level job satisfaction Estimates for Contextual Effects Earnings. We start with emphasizing that earnings may be a source of endogeneity in our regression, since unobservables determining job satisfaction may be correlated with earnings. As a precaution, we report two results: one with earnings as a regressor and one without earnings. To comply with the conventions in the income-comparisons literature, we construct our earnings variable, Income, as a dummy variable taking the value 1 if the worker earns greater than the median earnings in her reference group and 0 otherwise. We find that including this dummy variable into our regression does not alter the results significantly [see Table (4)]. We document that an individual worker s self-reported job satisfaction tends to go up with earnings, controling for social interactions effects. This is consistent with the findings in the income-comparisons literature, which suggest that job satisfaction depends on relative income comparisons [see, for example, Card, Mas, Moretti, and Saez (2010)]. The group-level results are surprisingly different. We find that the fraction of workers who earn more than 15 See Tumen (2012) for a theoretical discussion linking optimal group design to complementarities and increasing returns at the group level. 16

17 median wage earner in the reference group does not have a statistically significant effect on individual-level job satisfaction. This novel result suggests that the worker s job satisfaction level is determined by her own earnings position in the relevant reference group, not by the general fraction of high earners in the group. This result is the group-level analogue of the well-known Easterlin paradox [Easterlin (1974)]. Easterlin reports that, within a given country, individuals with higher income levels tend to report higher subjective well-being. But, at the international level, self-reported subjective well-being does not vary significantly with country-level per capita GDP. 16 We find a similar result at the group level. We show that workers with higher income levels are more likely to report higher job satisfaction, while there is no statistically significant relationship between job satisfaction and income at the group level. Gender and marital status. It is well-documented in the empirical literature that female workers are more likely to report higher job satisfaction. Our findings confirm this view; that is, the male dummy has a statistically significant negative sign. This finding is reversed at the group level though. More specifically, we find that individual workers tend to report greater job satisfaction in groups with a higher fraction of male workers. We have two dummy variables regarding the marital status. The first one marks whether the worker is married or not. The second dummy marks whether the individual worker is single or not, i.e., whether he/she has a partner or not, regardless of being married. We show that married workers and workers with a partner are more likely to report higher job satisfaction at the individual level. This is consistent with the main consensus in the literature. The group-level results are, again, different from the results at the individual level. We find that individual-level job satisfaction is lower in groups with a larger fraction of married workers, whereas there is no statistically significant relationship between job satisfaction and having a partner at the group level. Age. To account for the effect of age on job satisfaction, we include workers age as a linear 16 See Clark, Frijters, and Shields (2008) for an extensive review of the related literature. 17

18 in quadratic explanatory variable into our main regression equation. We show that older workers or, in the Mincerian language, workers with more labor market experience tend to report lower job satisfaction at the individual level. The sign of the quadratic term tells us that individual-level job satisfaction declines with age at an increasing pace. The grouplevel results are similar to the individual-level results. More specifically, we document that job satisfaction is lower in groups with higher average age and the decline becomes more significant as the average age goes up. Education. In line with the findings in the literature, we show that individual-level job satisfaction is higher for less-educated workers. In particular, we document that workers with college education and above are the least satisfied ones jobwise. Workers with no qualifications, on the other hand, are the most satisfied ones. The group-level results are surprisingly different. We find that self-reported job satisfaction is higher in groups with a larger fraction of higher-educated workers (i.e., workers with a graduate degree) and with a larger fraction of workers with A -level degrees. Signs of the remaining group-level educational dummies are statistically insignificant. Contractual status. There are two dummy variables describing the contractual status of workers: temporary contract and fixed-term contract. Working under a fixed-term contract is not a determinant of job satisfaction both at the individual and group levels. Working under a temporary contract, however, is a statistically significant determinant of individual-level job satisfaction. We show that workers under a temporary contract tend to report lower job satisfaction. Moreover, we find that individual-level job satisfaction is lower in groups with a larger fraction of temporary-contract workers. Employer status. We control for the employer status with two dummy variables: public sector worker and small employer. 17 Working in a small firm is positively related to individuallevel job satisfaction, while there is no statistically significant relationship at the group-level. Our findings for being a public sector worker are more interesting. The plain ordered-probit 17 Firms with size 200 or lower are classified as small. 18

19 results [see Table (2)] show that public sector workers are more satisfied than the private sector workers jobwise. When we control for social interactions, however, this result is reversed; that is, we find that individual-level job satisfaction tends to be lower for public sector workers. The same is true at the group level. In other words, we show that individual-level job satisfaction is lower in groups with a larger fraction of public sector workers. Union membership. Parallel to the findings in the literature, we show that being a union member is negatively related with job satisfaction at the individual level. We further show that the relationship between job satisfaction and union membership at the group level is statistically insignificant. Promotion opportunities. Workers who have access to greater promotion opportunities are more satisfied with their jobs, as expected. Similarly, we document that individual-level job satisfaction is larger in groups with a greater likelihood of promotion opportunities. Preference over working hours. Indicating a preference over working hours is negatively related to job satisfaction at the individual level. More specifically, we show that those who prefer to work more hours and those who prefer to work fewer hours both report lower job satisfaction. This maybe due to any discomfort related to the current working hours of the workers. We further show that individual-level job satisfaction is higher in groups with a higher fraction of workers indicating a preference to work fewer hours, while there is no statistically significant relationship between job satisfaction and preference to work more hours at the group level. Health status. We find that job satisfaction tends to be higher for those who report very good health and lower for those who report satisfactory health. The group level results are somewhat contradictory. In particular, we document that job satisfaction is higher in groups with a greater fraction of workers reporting satisfactory health, whereas there is no statistically significant relationship between job satisfaction and very good health at the group level. 19

20 5 Concluding Remarks In this paper, we manipulate the day of interview (also known as the the day-of-the-week) question in the BHPS data to test how individual-level job satisfaction is affected by grouplevel job satisfaction which is endogenous as well as contextual social characteristics in the worker population. We find that both endogenous and contextual social effects play crucial roles in the determination of individual-level job satisfaction. We present prima facie evidence that individual-level job satisfaction is higher in groups with higher average job satisfaction. We also document a rich set of findings regarding the (contextual) effects of group-level demographic, job-related, and employer-related characteristics on individual-level job satisfaction. To summarize our findings: we document that there are significant endogenous social effects in job satisfaction. This means that individual-level job satisfaction is likely to be higher in groups with higher average level of job satisfaction. As for the contextual social effects, we find that individual-level job satisfaction is higher in groups with a greater fraction of younger (or less experienced in the Mincerian language), male, and nonmarried workers. Moreover, we document that individual-level job satisfaction is greater in groups with a larger fraction of higher degree (i.e., graduate education) and A -level degree holders. Groups with greater promotion opportunities are more satisfied jobwise. We further report that individual-level job satisfaction is lower in groups with a greater fraction of public sector and temporary-contracted workers. Having satisfactory health and preference to work fewer hours are also positively correlated with individual-level job satisfaction. We report no statistically significant effect of group-level earnings on individual-level job satisfaction, although individual-level earnings have a positive effect on job satisfaction. Our findings related to the existence of endogenous social effects in job satisfaction are consistent with the main insights documented in the literature investigating social interactions and associated spillover externalities. Overall, we show that group-level job satisfaction represented as the mean outcome in the reference group has statistically significant effects on individual- 20

21 level job satisfaction. This suggests that individual-level and group-level outcomes, regarding job satisfaction, are complementary. In other words, workers in groups with greater job satisfaction levels are on average more satisfied jobwise. This has important policy implications, since it is well-known that job satisfaction is positively correlated with worker productivity and existence of positive spillover externalities at the group-level suggests increasing returns to policies fostering individual-level job satisfaction. 21

22 References Akay, A. and P. Martinsson (2009). Sundays are blue: Aren t they? The day-of-the-week effect on subjective well-being and socio-economic status. IZA Discussion Paper No Akerlof, G. A., A. K. Rose, and J. Yellen (1988). Job switching and job satisfaction in the US labor market. Brookings Papers on Economic Activity 2, Aradillas-Lopez, A. (2011). Nonparametric probability bounds for Nash equilibrium actions in a simultaneous discrete game. Quantitative Economics 2, Argyle, M. (2001). The Psychology of Happiness. London: Routledge. Arrow, K. J. and F. Hahn (1971). General Competitive Analysis. New York: Holden Day. Bala, V. and S. Goyal (2000). A non-cooperative model of network formation. Econometrica 68, Becker, G. S. and K. M. Murphy (2000). Social Economics: Market Behavior in a Social Environment. Cambridge, MA: Harvard University Press. Benabou, R. J. (1993). Workings of a city: Location, education, and production. Quarterly Journal of Economics 108, Bewley, T. (1981). A critique of Tiebouts theory of local public expenditures. Econometrica 49, Blume, L. E., W. A. Brock, S. N. Durlauf, and Y. M. Ioannides (2011). Identification of social interactions. In J. Benhabib, A. Bisin, and M. O. Jackson (Eds.), Handbook of Social Economics, pp New York, NY: Elsevier. Blume, L. E. and S. N. Durlauf (2001). The interactions-based approach to socioeconomic behavior. In S. N. Durlauf and H. P. Young (Eds.), Social Dynamics. Cambridge, MA: MIT Press. Brock, W. A. and S. N. Durlauf (2001a). Discrete choice with social interactions. Review of Economic Studies 68,

23 Brock, W. A. and S. N. Durlauf (2001b). Interactions-based models. In J. J. Heckman and E. E. Leamer (Eds.), Handbook of Econometrics, Volume 5, pp New York: North-Holland. Brock, W. A. and S. N. Durlauf (2002). A multinomial-choice model of neighborhood effects. American Economic Review 92, Card, D., A. Mas, E. Moretti, and E. Saez (2010). Inequality at work: The effect of peer salaries on job satisfaction. forthcoming, American Economic Review. Clark, A. E. (1997). Job satisfaction and gender: Why are women so happy at work? Labour Economics 4, Clark, A. E., P. Frijters, and M. Shields (2008). Relative income, happiness, and utility: An explanation for the Easterlin Paradox and other puzzles. Journal of Economic Literature 46, Clark, A. E., Y. Georgellis, and P. Sanfey (1998). Job satisfaction, wage change, and quits. Research in Labor Economics 17, Clark, A. E. and A. Oswald (1996). Satisfaction and comparison income. Journal of Public Economics 61, Cooper, R. and A. John (1988). Coordinating coordination failures in Keynesian models. Quarterly Journal of Economics 53, Easterlin, R. A. (1974). Does economic growth improve the human lot? In P. A. David and M. W. Reder (Eds.), Nations and Households in Economic Growth: Essays in Honor of Moses Abramovitz. New York, NY: Academic Press. Freeman, R. B. (1978). Job satisfaction as an economic variable. American Economic Review 68, Gabszewicz, J. J. and J.-F. Thisse (1996). Spatial competition and the location of firms. In R. Arnott (Ed.), Location Theory, pp Amsterdam: Harwood Academic. 23