$1.; .A .32.". . . , ,. . ., . . . .. r , . 3 i .. ..... LC. 3 z .{Luisr‘ , . . n. . Ln} : . I... . . | ‘13 2. . . . If xuapu 5.. 2a.: infiniflsv? . .; , I‘L rsfiumm. . fizzfieérnflhrm; a 352*“.181 . \l a £53.! .83. .31.. l . *3... 1.. an 3.. s. ‘d=\;‘x§ '- id! 1 .32 v: 1 .13.: .1 I! 31...!!!) h‘._.33\)hl§u|.z!l.i’ .. {In L unfit. 3.91:5!) .. .3! .. x. .. ...;....!.: ,. ‘ l. . . ‘ . ‘ . V . (a... a .1 ‘ \ .9322. .31.. 4......1 «9 11.....E.;....l. ‘. .131: Z Tjilvt‘l‘wyh _?66; This is to certify that the dissertation entitled EARNINGS OF SELF-EMPLOYED WORKERS AND PEER EFFECTS AMONG TEENAGERS presented by Daiji Kawaguchi has been accepted towards fulfillment of the requirements for Ph.D. degree in ECOHOIIIiCS Major professor Date E/Z’é/ZOOZ‘ MS U is an Affirmative Action/Equal Opportunity Institution 0-12771 ¥ i .UBRARY Michigan State University PLACE IN RETURN BOX to remove this checkout from your record. TO AVOID FINES return on or before date due. MAY BE RECALLED with earlier due date if requested. DATE DUE DATE DUE DATE DUE 6/01 c:/ClRC/DateDue.p65-p. 15 EARNINGS OF SELF-EMPLOYED WORKERS AND PEER. EFFECTS A.\IONG TEENAGERS By Daiji Kawaguchi A DISSERTATION Submitted to Michigan State University in partial f1.11fillment of the requirements for the degree of DOCTOR OF PHILOSOPHY Department of Economics 2002 ABSTRACT EARNINGS OF SELF-EMPLOYED WORKERS AND PEER. EFFECTS AMONG TEENAGERS Bv v Daiji Kawaguchi This dissertation contains three essays in applied microeconoinics. Chapter 1 revisits the empirical results of Lazear and Moore [Edward Lazear and John Moore (1984) “Incentive, Productivity, and Contract” Quarterly Journal of Eco- nomics 99]. That paper found that empirical experierice-earnings profiles were flatter for self-emplrwed workers and argued that. this supported the Lazear contract theory that claim firms use life-cycle backloaded payment systems to work around principal-agent problems between firms and workers. This chapter reproduces the Lazear and Moore result on more modern data, but argues for an alternative interpretation. In particular, this chapter argues that. self—employed workers face more wage variation but enjoy a higher return for human capital. A model based on these assumptions can produce flat— ter experience-earlrings profile. since selflemployed workers start their career with more human capital and due to opportunity cost, they invest less in human capital on the job. The chapter develops implications of the model not found in the Lazear contract model and concludes by developing support for these implications. Chapter 2 attempts to explain the lower earnings among self-employed workers found by Hamilton [Barton Hamilton (2000) "Does entrepreneurs]rip pay? An Empirical Anal- ysis of the Return of Self-Employment” Journal of Political Economy 108]. That paper found 20% lower ez-u‘nings of self-employed workers with 10 years of business tenure than comparable salaried workers with 10 years of job tenure. This difference in earn- ings can in principal be explained by the cmnpensating wage differential theory when self—en’iployed jobs have attractive uon-numetary aspects. Using the National Longi- tudinal Survey Youth 79 (NLSY'TQ), this chapter tests whether self-employment is as- sociated with higher global job satisfaction. By looking at changes in job satisfaction for individuals over time. I overcome the difficulty of interpreting differences in sub jec- tive job satisfaction scores across individuals that cross—sectional analysis would require. Using my estimates, I calculate the monetary value of the non-monetary aspects of self- employment and find that. one dollar earned while a self-employed worker is equivalent to as much as three to four dollars earned as a salary or wage worker. Although the valuation is surprisingly high but the the, directimi of the estimate is consistent with the compensating wage differential hypothesis. Although job satisfaction is a partial component of workers" total utility. the value of self-employment in terms of job sat- isfaction is sufficiently high to support the compensating differential hypothesis as an explanation for lower earnings among self-employed workers. I also evaluate several other explanations for the surprisingly high valuation of self-employment. Chapter 3 attempts to estimate peer effects on substance usage among teenagers. This chapter first summarizes the problems in the identification of peer effects. The existence of unobserved characteristics of individuals and endogenous sorting into refer- ence groups based on unobserved characteristics causes problems in the identification of peer effects. The solutions for this problem are: 1. To control “imobservaliile” through including plenty of explanatory variables using rich data set. or using sibling method to difference out nnol')serval)le. 2. To use natural experimental situation in which reference group is assigned randomly. 3. To use economic theory to get a prediction that arises only from peer effect. but. not from contextual or correlated effect. In this (’rl‘iapter, the method 1 was taken. Significant peer effects were found on substance usage among teenagers. ACKNOWLEDGEMENTS I would first like to thank Professor David Neumark. the chairperson of my disserta- tion committee. Before taking his class, I was womlering if I can study American labor markets as well as American researclmrs because of my lack of experience in American society. When I expressed this concern in a class presentation, he just recommended me to look at data which is pulflicly avail-able. This open minded attitude of him toward the study of American labor market largely removed my concern and, consequently, I used American data for all three chapters in this dissertation to address several problems in labor economics. He gave me critical connnents on my marmscripts and taught me how to write papers. In addition, working as his research assistant was a great opportunity of apprenticeship. He taught me lessons through "teaching by doing” that I am very grate- ful. Professor Jeff Biddle kindly participate in my dissertation commitee and gave me well thought comments on my manuscripts in very timely manner. He has been always generous to take time to address my concerns. in particular, when Professor Nuemark was on leave in the fourth year of my doctoral study. Professor Jeffrey “fooldridge also kindly took part in my dissertation connnittee and took his time to discuss econometric issues in my dissertation. His distinguished approach to econmnetrics enlightened me and sparked my interest in applied econometrics. The third chapter of this dissertation is the outgrowth of the term paper that I sul’n‘nitted for his applied econometrics class. Without help of three members of my dissertation committee, I could never finish this dissertation. Professor Charles Hadlock from Finance Department. also kindly took part in the dissertation defense as an outside committee member. I am also grateful to professors who taught classes and participated in my seminar presentations. In particular, I would like to thank Professor Gerhard Glomm, Professor John Strauss, and Professor John Godderris. Professor Glomm taught me neoclassical economics starting from the definition of “Economy.” I especially learned a lot. through writing a paper with him and Far-undo Sepulveda. Professm‘ Strauss and Professor God- iv derris regularly attended my seminar talk and gave me. comments and encouragements. During four and half years of my doctoral study, I learned as much from friends as from professors. In particular. the conversation with Scott Adams, Linda Bailey, Ali Berker, Rehim Kilic, Jai—Min Lee. Zhehui Luo, Facundo Sepulveda. Jeongseok Song, Chien-Ho Wan have stimulated my interest in many research fields of economics. I am also very grateful to all their friendship off campus. I deeply appreciate all participants of labor lunch workshop who kindly listen to my presentation and comment on my work. All graduate students from Japan in both economics and agricultural economics department. in particular, Hirokatsu Asano, Tatsushi Adachi, Kei Kajisa, Yoko Kijima, Horoki Nakayama. Katsushi Okada. lx'yota Yamada. and Takashi Yamano have made me feel at home in East Lansing. My wife, Yukimi. has patiently understood and supported my study at Michigan State mostly from Japan. The time I could live with her in Ann Arbor for one year was the most delightful time during my stay in Michigan. I am very much indebted to my parents for their support in various ways. My parents experience as self-employed workers largely motivated my research interest in self-employment. I am very grateful to Yukimi and my parents for their unconditional love to me. I hope my receipt, of PhD. degree will give great relief for them as well as for me. \f Contents 1 Human capital accumulation of salaried and self-employed workers 1 1.1 Introduction .................................. 1 1.2 Replication of Lazear and Moore (198lf-employed workers face the larger risk, we are interested in if (p = 0. The model that is actually estimated is In 11'), = Xm31 + 13-3.», + SMXI': — X);33 + c, + c” + (31 + bis” + 21,}. (1.5) Applying the fixed effects transforination, we obtain In 113;, —— Iii—w,- = (Xi, —- X1331 + 32(5): —— 5,) + (8”le —fi).;33 + cit + b,(s,;, —.§,,:) +11” (1.6) In this situation. the fixed effects estimator is not a consistent estimator since j.)li1n."32 = .32 + b, wl‘iere b : 13(1),). Thus the fixed effects estimator of :32 estimates the lower bound of ,3? given I) < 0. Another possible. measurement. error arises from the differential concept. on the in- come between SW’ workers and SE workers. SE workers may report return for physical capital of their own business as their wages: they also may subtract the cost of physical investment in their own business from their wages. It is likely that SE workers invest in physical capital when they start their businesses and collect the return later. Then the wages of short. temu'ed SE workers are understated and the wages of long tenured SE workers are overstated. Due to this measurement error, the return to tenure among SE workers may be overestimated. Regardless of this possilnlity of upward bias, the estimated return to tenure among SE workers are almost zero. The conditional variam':e of measurement error may also depend on self—employn‘rent status through the effect of bi. Define the error term in (1.6) as he 2 e“ +u,~, +1),(s,t —.§,—). Then Elllfr X7381] = (1+ 05‘”)fo + (7?, + (52'! —” giyElbflxri‘ 8'3” A .—l “J V q I The last term tells us that. the job status changer tends to have a larger variance. As- smning E[I)?|X,:. 5,] = (7,3). regressing the residual of fixed effects wage equation on Si, and (st, — .97.)"2 renders consistent estimators of e) and 03. The OLS and fixed effects estimates appear in column (5) and (6) of Table 1-1. Although the estimate of go dimin— ishes slightly. it is still large and statistically significant. Therefore, we still conclude that. being a self-employed worker is riskier than being a wage-salary worker. 1.3 The model Let us suppose that. each worker lives for two periods and that each worker is endowed with one unit of time for each period. Each worker knows his ability in the first period. Each worker has the following preference with constant absolute risk aversion: [/11 = " (“Pl-'iifu'n + “321] (1-8) where on is the degree of absolute risk aversion of worker 1'. and Hf” is the wage offer for worker i at time t. The wage offer depends on job choice, meaning whether the worker is self-enmloyed or salaried worker. The wage offer for job j is “’in =()J'(1— n.,:,)/1,-, + 6’)”, €in N 1V“). Uf)f01']= SE, SH], (1.9) where 11., is the human capital of worker i. at period t, 11., E [0.1] is the portion of time devoted to the human capital accumulation by the worker i at time t. The initial human capital hi1 is given as an endowment for each worker and includes human capital accumulated through education and innate ability.7 The human capital for both periods is assumed to be general across the jobs. The parameter bi: which is exogenously given by the labor market for the workers. is the unit price of human capital in job j. The random variable c,”- is a shock to the wage. This model assumes e,“- is independently distributed across individual, time. and two jobs. Taking expectation of the life-time ‘While recognizing the emlogeneity of the educational decision. I have treated this as given since the main interest of this analysis is on-the-job human capital accumulation. utility and using the ordinal property of utility function, ELY =()J'(1— Ili1)}l,‘1—(’7,/2)0’j+()j(1—n,'2)}lig ‘— (Ali/2)”? (1.10) J is obtained}8 Each worker has access to the following human capital accumulation teclimiilogy: hi? :1?” + (5(Hl'1hl'1)o.0 E (0.1), (1.11) where 6 is efficiency of human capital investment on human capital accumulz-ition.9 The parameter 0 represents a workers learning ability. which is assumed to be identical for all workers. This technology shows that the worker with higher human capital is more productive in the production of additional human capital, but the effect diminishes, since a E (0.1). Two assumptions are made that distinguish SW and SE workers. . I .) . , . . AssumpMm 1 (73).); > (raw. 1.e.. the wages of SE workers are more volatile than those of 8“" workers. The empirical evidence supports this assumption as seen in the previous section. Assumption. 2 by; > (my. i.e., the return for human capital is higher for self-employed workers. This assumption is justified by the higher returns to education among SE workers than SW workers in previous study that utilize Census data.10 For example, Fairlie and Meyer (1990) found 0.90 as the return to education among SE workers while they found 0.59 among S“? workers. The data used in this study does not indicate higher return to education. though. Under these assumptions, each worker maximizes his lifetime expected utility (1.10) . . ') . . by choosing nil and a. career path ( {J};:1) under the constraint of human capital accu- 6See Appendix for derivation. 9This functional form of human capital accumulation is standard in the literature. See llecknian (1976). U’See Borjas and Bronars (1989) and Fairlie and Meyer (1996). In particular Fairlie and Meyer (1990) used 1990 Census Public Use Blicrodata 5-percent sample that contains 11881 SE sample and 100-314 SW sample. mulation technology (1.11).” Since indirect. utility from the career path of SE—SW’ is always dominated by the indirect utility from other career 1.)athes. we should consider three possible career paths of SE - SE. S\-\" - SW, and 8“" - SE. The optimal human capital investment time, n”, is (in I h—l. for '01) stayers. I)” Z (I) ) ll 1 _1 J ~ (1.12) (“fin )1 ~ (do)l “h,1 , for job changers These solutions show the human capital investment time decreases in the level of initial - ') human capital.” By substituting the optimal 11,1 in the objective function of each career path, we obtain the following indirect utility functions for each career path for each individual i. l . -4 _e_ _1_ . . l‘j—jU’il-AH) 212(0le —",','(Tj-2+bj()‘ ” (01"0 —(.11 “) for] =3 SE.’SH. (1.13) l'su'—s1«:(hz‘1~ 7:) = (bsu' +bs‘F31hil " (Vi/21(Ui'u' + (7.31?) '_” ; n +b;,,«~b;.,;6n—‘n(afi — 0+») (1.14) These expressions tell us that the choice of career path depends on each worker's level of initial human capital and the degree of risk aversion. The relationship between lifetime utility for each career path and initial human capital is graphed in Figure 1-1 given the degree of risk aversion. This graph shows the lifetime utility of being a SE worker is higher than being a SW" worker for the high human capital worker. In addition. the graph shows that. the worker with a “medium" level of human capital switches jobs in the middle of his career. The worker with higher initial human capital selects self-em})loyment. This selection affect the wage growth of SE and SW workers. The "average" wage growth for each 11As a result of optimization. 71,2 = 0 is trivially chosen. 1“’There are two factors related to the initial human capital level and human capital investnu-‘nt. First. from (1.11) the worker with higher human capital is more 1')roductive in human capital accunnilation. Hmvever. this effect is diminishing becmrse the term (”uh”) is exponentiated by a E (0.1). Second. the worker with higher human capital pays more opportunity costs of human capital investment, which is bjnnhn. The marginal benefit of investment diminishes in hn but the marginal cost is constant in lid: thus the worker with high [1,1 chooses lower ”,1. The convexity of human capital production function is the crucial assumption to derive this result. 10 career path is, 1“ Eu,” h” +61/(1—a)na/(1—0) . r 9]-]- — EMU —- hi1_ (SI/(l—HJQI/'(1~n) for j —— SE,SH, (1.15) . ~ _ Ewes»: _ bse ha + (ll/(1‘0)((bSE/bsn')0')a/(1_0) . (1 16) 9511—.8E —EU‘15-'u' b3”, h” _ (51/(1—0)((bSE/bsw)a)l/(l—Q)- . If the initial level of human capital is given, the wage profiles of the SE and the SW workers are identical and it ('lecreases as hi1 increases and converges to one. However, what. we observe is (£575-51; < ggu-_5u- because the SE worker’s 12,-1 is higher than the SW worker’s. Thus the observed difference between wage profiles is the result of the workers" heterogeneity in the initial level of human capital. 1.4 Supporting evidence of the model 1.4.1 SE workers have higher human capital The theory discussed in the previous section predicts that the worker with higher human capital selects self-em})loyment. With respect. to the observable characteristics. several studies report that the worker with high education is more likely to be self—employed.14 It is a stylized fact that the worker who has a self-employed father is more likely to be a SE worker, even after controlling for inlierita.nce.15 In particular, Dunn and Holtz- Eakin (2000) emphasize the importance of the intergenerational transmission of human capital rather than the mitigated liquidity constraint to explain this finding, since they found a very large effect of the parent's self-employment status on the son's selection into self-ernployment even after controlling the amount. of the parents” assets. They also found that the son of a successful self-employed worker is likely to be self-employed. From this finding. they conclude that the transmission of human capital is the important. channel to explain the i1itergenerational correlation of self-employnient status. Although 13Although the expected value of a ratio is not a ratio of expected values. this measure gives us a rough idea. 14See Borjas and Bronars (1989) and Evans and Leighton (1989) 1“SSee Lindh and ()hlsson (1.996) Blanchflowm‘ and Oswald (1998). ”out and Rosen (1999) and Dunn and Holtz-Eakin (2000). By controlling inheritz‘ince. the researchers try to partial out the effect of liquidity constraint. ll their findings may imply the transmission of human capital that is specific to the SE, the theoretical discussion made in this paper still holds since the worker with "any kind of” higher human capital experiences lower wage growth. These results support the prediction of the model presented here: Self-employed workers experience lower wage growth because of higher level of initial human capital. 1.4.2 Less human capital accumulation among SE workers The model predicts less human capital accumulation among SE workers. This predic- tion can be directly tested through the comparison of recorded human capital investment behavior by SE and SW workers. School attendance or participation to training while working are good measure of human capital investment. The i’LSY79 records the par- ticipation to training in a consistent way after 1988 survey, thus analysis sample in this subsection is restrictml after 1988. This sample restriction reduced the sample size to 17825 for school enrollment analysis and 17818 for training participation analysis. This restriction also largely excludes college students who have part-time job since respon- dents are at least age ‘21 in 1988. Firstly, school enrollment while working is analyzed. Table 1-2 tabulates any school enrollment since last interview by the types of job. While 9 percent of SW workers enrolled in school between last and current interview, only less than 6 percent of SE workers enrolled. To control the difference in the characteristics of SW and SE workers, school enrollment was regressed on self-enipltwment dummy as well as observed char- acteristics by OLS. Fixed effects linear probability model was also estimated to deal with individual hcterogeneity. Both ()LS and FE results, which appear in Table 1-3, indicate that SE workers are about. 2 percentage points less likely to enroll in school while average enrollment rate is about 8.7 percent. Secondly, participation to training is analyzed. The participation to training is not straiglitforwardly cmnparable between SE and SW workers since some training that takes place in SE sector may not be recognized just because it does not take the form 12 of formal training. To work around this proli)lem, Table 1-4 Panel A breaks the partici— pation to training into types of training that workers participate. On-the—Job—Training ("Apprenticeship Program," “Formal Company Training Run by Employer or Military Training" or “Seminars or Training Programs at Work Not Run by Employer”) may not be conmaralfle between SE and SW workers since those training programs are less likely to exist. in SE sector in formal fashion and consequently actual participation to comparable inftn'n‘ial In'ograms by SE workers may be Inisclassified as non participa- tion. 011 the other hand. the participation to Off-the—Jol) Training (“Business School,” “Vocational or Technical Institute.” "Correspondence Course,” "Seminars or Training Programs outside of \\-"()rk" and “Vbcational Rehabilitation Center”) is comparable be- tween SE and SW workers since the definition of these training programs are presumably identical across SE and SW workers. Table 1-1 Panel A shows that 18.91 percent of SW workers participate in training while 9.21 percent of SE workers participate in training. Restricting our interest to Off-the-Job Training (Off-JT hereafter) renders the same picture, although the difference is largely reduced. While 6.20 percent of SW work- ers participate in Off-JT, 5.00 1_)ercent of SE workers participate in Off-JT. To control the difference in observed characteristics between SE and SW workers, liner probability models with and without. fixed effects are estimated. The OLS result. that appears in Column (1) of Table 1-5 indicates that SE workers are 9.7 percent points less likely to participate in training. while FE result that appears in Column (2) of Table 1-5 indi- cates that SE workers are 5.0 percent points less likely to participate in training. As for Off—JT, OLS result. that appears in Column (3) of Table 1-5 indicates that SE workers are 1.2 percent points less likely to participate in Off-JT while FE result (Column (1) of Table 1-5) indicates 1.0 percent points reduction. All of the results above indicate that SE workers are less likely to participate in training and accordingly invest less in their human capital on the job. The model also predicts that 8“" workers who become SE in the second period invest more in their general human capital in the first period because of lower opportunity 13 cost of human capital investment and higher return to human capital. To test this theoretical predicticm. using SW workers as sample, the participation to training is regressed on future SE dummy and other controls by OLS.l6 The result of estimation appears in Column (5) in Table. 1—5 but. this result indicates that future SE workers are less likely to participate in training in general. This result does not necessarily contradict with the theory since SW" workers who become SE in the future may invest. less in firm specific human capital. Thus it is important. to disentangle investment in general human capital from investment in firm specific human capital. Ideally, general human capital investment should be regressed on future SE dummy and other controls. As an attempt. to implement this idea. the participation to Off-JT or training whose cost is paid by workers or government are regressed on future SE dummy and controls. In particular. the infm‘ination who heard the cost of training is important. since Becker’s traditional theory predicts that. training that endows workers with general human capital is paid by workers while the cost of investment in firm specific human capital is shared by both workers and firms. Although training whose direct cost is paid by employer might be excrrtually paid by workers through lower wage, training whose direct cost, is paid by employee or government bounds the minimal set of training that presumably endows workers with general human capital. The results of regressions of Off-JT and own/goverment paid training participation on future SE dummy and other variables appear in Column (6) and (7) of Table 1—5. The participation to Off-JT does not differ practically among SW workers who stay in firm and who become SE in the future. However. future SE workers are 1.1 percent more likely to participate in the training whose cost is paid by his / her own or government. Considering only 3.3 percent of SW’ workers participate in own / government paid training. the difference is large. This result imply that prospective SE workers invests more in general (portable) human capital while they are SW workers. 1“Future SE dummy varies within individual for those who switch to SE and switch back to SW but no prospect to be SE again. Since identification based on those observations are. weak. fixed effects estimation is not illl})lt'lll(_‘llt(’(l. 1 1 1.4.3 Both Winner and loser select SE? The model predicts a very simple selection rule: The worker with high human capital selects SE. A sensible criticism for this prediction is that there are two kinds of SE workers. The first. kind is an eligible entrepreneur and the second kind is the SE worker who is not. qualified to work for a. firm and is forced to work by himself. If this story describes the real world. it is not surprising that the latter group experiences less wage growth since the less eligible worker has less learning ability and experiences less wage growth. To examine this possibility. I studied the distribution of ability among self-eiiipltjiyed and wage—salary workers. If there is a. “two tail selection rule" among SE workers, we should find bimodal distribution of ability among SE workers that have two peaks at. high and low ability. As a. proxy for the ability I used the AF QT89 (Armed Force Qualifying Test) score that is contained in NLSY 79. The result. of the kernel density estimation of the. distribution of test scores for the SE and SW workers appears in Figure 1—2. Comparing the two distributions. we find bimodal distrilmtions among SW workers and SE workers. However SW workers have peaks on high sores and low scores but SE workers have perks on high scores and medium scores . This evidence shows that “two tail selection rule" is more likely among 8“" workers than among SE workers. 1.5 Conclusion In this paper. the human capital accumulation by self-employed (SE) and salaried and wage (SW) workers was ans-ilyzed. Under the assumptions that the wages of SE workers are more volatile than salaried workers and the wages of SE workers more sharply reflect their human capital. SE workers invest less in their human capital because of their higher initial human capital. This ('lifference in human capital investment behavior results in a flatter wage profile for SE than SW workers. This theory was indirectly supported by the empirical facts about self-employed workers. In particular. the empirical evidence 15 shows that the SW workers with future SE prospect. accumulate more general human capital on the job and this evidence is consistent with the model developed in this paper. The model shows that the Self-(‘1Il[)l()_\'(.?('l workers are not necessarily a good "control” group to test the Lazear contract. since not only does the incentive effect of the Lazear contract produce the steeper wage profile of salaried workers. but the difference in human capital investment has this effect as well. This conclusion does not. deny the existence of the Lazear contract. nor the results produced by Lazear and Moore (1984). However, simply attributing the difference of wage profiles to the incentive effect of the Lazear contract may overt-istimate its importance. The full-blown test of the theory introduced in this paper is left. for future research. Appendix The derivation of (1.10) is as follows: Eff” : //—expl—7,-(Ir,-1 + 11'.2)](1F(c',-1}-)(IG'(c*i-_)J-) = ’/9Xl)l7i(f’J‘/H1(1“’ln)+ €iji)l(/F((’U‘1) -/exp[e,i(bJ/2,~2(1— r1i2)+ eij2)](lG(c,-jg) = — CX])[—‘},;(f)J-(1 — 11.1)12” — (“H/2M; +bjhi‘2 —' (’i‘i/‘Zlf’fll. The independence of error terms across periods derives the second line and the property 0f 10g normal distribution produces the third line. When ln.r ~ .n\¥'(‘172..s‘2). it. is known that. Er = exp(m + (fl/‘2).s2). In our case, e = 1111' ~ ;\’(Ilz..52) thus Eexp(e) 2 E1: E‘XD( m + (1/2)82). The ordinal property of utility fimction results in (1.10). 16 Chapter 2 Compensating wage differentials among self-employed workers: Evidence from job satisfaction index 2. 1 Introduction Self-employed wcn‘kers comprised 10.5% of the total US. workforce in March 1996.1 Despite this large share of self-employed workers. self-employment has not attracted much attention among labor economists until recently. and the workings of the labor market among self-employed workers are still largely unknown. One of the remaining puzzles about self-employed workers is their lower earnings, which this paper attempts to explain with compensating wage differential theory. The compensating differential theory predicts l(_)wer earnings among self-en'iployed workers when non-earnings aspects of self-employed job positively affect workers” utility. This paper directly tests the theory using job satisfaction scores available in the National Longitudinal Survey Youth 79 (NLSY79). Several studies report lower earnings among self-employed workers as compared with their salaried and wage-earning counterparts.2 For example. Hamilton (2000) found lFor recent trends of self—employment in the U.S.. refer to Mauser and Picot (1999). 2The analysis in the previous chapter described self-employed sector as the sector where the return to human capital is higher and consequently workers with higher human capital self-select into self- en‘iployment. In total. self-employed sector was described as better paid sector. At first glance, this chapter and the previous chapter may look contradicting. but if the utility from non-monetary aspects of self-employed job is introduced in the model of the previous chapter. then discussion in the previous chapter still carry over but we may observe lower earnings among self-employed wm'kers. This can be 17 that self-employed workers earn less than salary / wage workers with similar. observable qualifications. using several measures of self-employed workers’ earnings available in the Survey of Income and Program Participation (SIPP) 1984 panel. These lower earnings are mainly due to a lower growth in earnings among self—employed workers. According to Hamilton’s findings. on average. self-employed workers with 10 years of business tenure earn 19% less than salary/ wage workers with the same amount of work experience. Lower earnings growth among self—employed workers was also found by Lazear and Moore (1984). Krashinski (2000) found 10 to 30% lower median earnings among self- employed workers. except for college graduates. after controlling for observable workers’ characteristics using CPS files over the period of 1979 - 1992. Carrington. McCue. and Pierce (1996) also found similar earning differences using the Current Population Survey (CPS) March files between 1967 and 1992. These observed lower median and mean earnings among self-employed workers are rather puzzling. llC)VV'(‘V€l‘. considering the self—employed workers labor income risk as reported by Carrington. McCue. and Pierce (1996).3 They found that self-employed workers labor earnings are three times as sensitive to macro aggregates as salary / wage workers’ and concluded that the labor earnings of self—employed workers are much more pro—cyclical to the business cycle. In addition. self—employed workers tend to use their own assets as capital for their own businesses due to liquidity constraints (Evans and Jovanovic ( 1989)) and. as a result. their labor earnings and asset income tend to co—move. This co—movement nrakes it difficult for self-employed workers to insure their future consumption, compared with wage / salary workers who can insure this by saving in a safe asset. In addition. I\'Ioskowitz and Vissing-Jorgensen (2001) showed self-employed workers tend to invest a large portion of their assets in their own businesses. As a. result. easily seen from Figure 1-1. If noanonetary aspects of self-employment generate utility. then indirect utility of SE - SE or SW - SE career from monetary income shifts down while total indirect utility stay constant. This shift may create lower average earnings among self-employed workers. 3Labor income risk among self—en‘iployed workers is intuitively appealing. and consequently several theoretical papers employ it as an assumption that characterizes self-employed jobs. See Kililstrom and Laffont (1979) and Kanbur (1982) for example. 18 self-employed workers" portfolios are riskier than those of salary/ wage workers, whose assets are invested in more diversified funds. Despite this risk. the average return on portfolios held by self-employed workers is almost equivalent to the average return on portfolios held by salary/ wage workers. To compensate for this total income risk. at least at the first glance. the average of earnings for self-employed workers should be higher than that for salary / wage workers. In addition to income risks, self-employed workers enjoy fewer fringe benefits. such as employer-provided health insurance. than salary / wage workers. as pointed out. by Hamilton (2000). Considering the negative aspects of self-emplovmeiit. the lower earnings of self-employed workers is a puzzle. \Vorkers' negative self—selection into self—employment is one possible explanation for these observed lower earnings. If workers with negative unobserved characteristics self- select into self-enmloyrnent. the lower earnings observed among self-employed workers may be due to these negative traits. To evaluate this possibility, Hamilton (2000) compared the earnings of two groups of salary / wage workers during a two year period; the first group consisted of workers who became self-employed workers in the second year, and the second group consisted of workers who continued to be salary/ wage workers. He did not find any significant (Llifference in salary / wage earnings for these two groups and concluded that self-selection does not explain lower earnings among self-employed workers. Krashinski (2000) did the same exercise using matched CPS data for 20 years and found no evidence 0 positive or negative selection into self-employment.“1 Borjas and Bronars (1989) even found positive self-selection into self-employment among white ”13188. using a Heckman-style. self-selection correction method.‘5 To summarize. previous reSearch shows that negative self-selection into self-employment does not explain the .___4\ I'{I‘ashinski (2000) used this finding as evidence that self-employed workers are a good control group for testing the institutional hypothesis to explain wage inequality during the 1980s and 90s among salary / wage workers because self-employed workers are relatively free from institutional factors Snell.) as minimum wage restrictions and labor union involvement. QThe SMSA-level aggregate labor market conditions are used as instruments in the first. stage se- ecti011 equation. The variables are unemployment rate. population growth rate. crime rate. the level Pf local government expenditurt—x and the mean of income and education level. These varial')les are dSSimred to affect the labor market mobility but do not influence the determination of earnings. 19 lower earnings of self-employed workers. Hamilton (2000) speculated the compensating differential as an alternative expla- nation. He claimed that self-employed workers enjoy non-earnings aspects of self- employment. such as being their own boss. and. accordingly. they accept lower earn- ings. although his claim was not. supported by direct empirical evidence. In one sig- nificant study. Blanchflower and Oswald (1998) tested compensating differentials for self-employed workers using the job/ life satisfaction variable available in British Na- tional Child Development Survey. Using cross section data from 1981 and 1991. they found that self-employed workers are more satisfied with their job/life. than salary / wage wm‘kers. However. as authors admitted in their paper. there is a possibility that "self- employed people may be intrinsically more optimistic.” and higher job/life satisfaction among self-employed workers might be due to intrinsic characteristics of self-employed workers. Several psychological studies. in fact. have revealed that. people with positive attitudes toward life are more likely to be self-employed.6 This problem arises from the interpersonal comparison aspect of the job / life satisfaction score that. is determined by subjective perception. This problem can be resolved by considering the change of job satisfaction associated with changing jobs. because these sz-itisfaction scores are cour- pared within individuals. This possibility. however. cannot be explored without using panel data. In addition. contrary to the findings in Blanchflower and Oswald (1998). Clark and Oswald (1991). using a. medical measure of psychiatric health. found that self- employed workers are. more highly stressed than salary / wage workers. Considering the risks that self-employml workers face, this result is not surprising. Thus, the evidence for the compensating differential among self-employed workers found in Blanchflower and Oswald (1998) is very informative but it does not decisively support the compensating differential hypothesis. This paper attempts to overcome the limitation of Blanchflower and Oswald (1998) by using panel data with a subjective job satisfaction measure (Na- tional Longitudinal Survey Youth 79). and. in addition. it attempts to calibrate the “See Brockhaus and llorwitz (1986) for review of the literature. 20 monetary value of self—employmelit status. The rest of this paper is organized as follows: Section ‘2 overviews the use of job satisfaction measures in economics. Section 3 briefly discusses our data and confirms the lower earnings of self-etripltiiyed workers. Section 4 discusses job satisfaction scores in the data and implements a descriptive analysis. Section 5 describes a simple model of the compensating differential among self-employed workers and estimates the param- eters of the model. Section 6 extends the analysis in Section 5. relaxing the imposed assumptions. Section 7 provides a. summary and conclusion. 2.2 How Economists Have Used Job Satisfaction Measures Economists often hesitate to use subjective job-satisfaction measures because linking job satisfaction measures with underlying utility is thought to be (.lifficult.7 In the empirical literature. howevm‘. labor economists have made significant efftn‘ts to incorporate job satisfaction n‘ieasures into economic analyses of labor market outcomes. There are roughly two ways in which job satisfaction is used in economic analyses of labor markets. The first uses job satisfaction scores as an independent variable to examine the. effect of job satisfaction on economic outcomes. Freeman (1978) showed that job (dis- )satisfaction predicts workers’ job quitting behavior fairly well. even after controlling worker and job characteristics including their wages. C arrington. McCue. and Pierce (1996) and Clark (2001) obtained similar results using German and British data re- spectively. These findings establish that job satisfaction is a. very informative economic variable. The method that is going to be applied to the data. set in this study will help determine whether job satisfaction is a reliable economic variable. The results of the analysis will appear in data section of this paper. ‘Job satisfaction is regarded as one of most. important concepts by industrial and organizational psychologists. and textbooks in the field typically devote an entire chapter to job satisfaction and its effect on job performance. For example. see McCormick and Ilgen (1985) and Siegel and Lane (1987). ‘21 The second type of studies has analyzed the determination of job satisfaction. given that the job satisfaction variable is a reliable economic variable. The job satisfaction variable has been widely used to examine the effect of unionism on job satisfaction (Borjas (1979). Leigh (1979) and Bender and Sloane (1998)). Somewhat surprisingly, they typically found lower job satisfaction among union as compared with non-union workers. They explained this counter—intuitive finding as an evidence of an exit-voice mechanism of labor unions: even though workers are not happy with their jobs. they do not quit because their ”voice" is heard thrtiiugh labor union. Several studies have used job satisfz-tction scores to test the relative income concern ll)'}'.)()lll(".SlS; job satisfaction is not only determined by individual earnings. but also by relative position of the earnings compared with workers who share similar characteristics (Hamermesh (1977). Clark and Oswald (1996). Hamermesh (2001)). All of these. studies found evidence supporting the relative income concern hypothesis. This review of literature shows how the use of job satisfaction in economics has extended the area on which economics can shed light. 2.3 Data and Lower Earnings among Self-Employed Workers The National Longitudinal Survey Youth (NLSY79) is used in this study. The analysis sample consists of observations between 1985 and 1998 that was restricted to white male in order to be consistent with previous studies (Hamilton (2000), Krashinski (2000)). Iru’lividuals who work for money and are out of school are included in the sample. Individuals are dropped if their job classifications are unknown. The construction of the analysis sample is tabulated in Table '2 - 1. As the first. step of the analysis. the lower earnings of self—employed workers. which have been observed in previous studies. are replicated with this data. The. following wage equation is estimated to see the earnings (,lifferentials between self-employed and salary / wage workers. lll (1‘); = .30 + I113} 'i‘ {fgb'fffn + Selfulfitjfg + C,‘ + ()1?! , (2.1) where 21)., is hourly rate of earnings.t4 self,-, is the dummy that. indicates self-employed. 17,-, is the vector of human capital and demographic variables. c.- is unobserved indi- vidual heterogeneity. and e,-, is idiosyncratic error that satisfies E [eulself a. '17“. Ci] = 0. The model is estin‘iated through OLS assuming [(.',-|self,-,.:1:,-,] = 0. and this assump- tion ensures that self selection into self—employment does not occur on the basis of unobserved characteristics. \‘Vhen the assumption [c.jsel f it. In] = 0 is violated through self-selection based on unobservables. the OLS estimator is biased. To deal with this possibility. the model is also estimated through a fixed effects estimation. The fixed effects estimator is unbiased if e,-, is strictly exogenous (i.e. E [eitlsel f.. 27,-, 0,5] = 0. where self.- = [self.1.....scl f.-T].;r. = [r.1......r:.T]); thus self-selection into self-employment based on time—ccmstant unobserved characteristics is allowed. The differences in earnings between salary / wage and self—employed workers are eval- uated at 10 years of job market experience and 10 years of job (business) tenure.9 The point. of evaluation is important. since the life-cycle earnings profile is much flatter among self-employed workers as pointed out in Lazear and Moore (1984). The results of estimation and the estimated difference of earnings appear in Table ‘2 - 2. Both the results of OLS and the fixed effects estimation show that the earnings 8Hourly rate of pay is constructed by the Center for Human Resource Research (CHRR) based on respondents usual earnings (inclusive of tips. overtime. and bonuses but before deductions). CHRR requests wages/ salaries / tips income and business/firm income along with other income categories from both wage/salary workers and self-employed workers. Thus there is less concern that self—employed workers earnings contains capital income. in addition to labor income. which is the main concern for the measurement of income among self-employed when CPS is used. Therefore the measurement of earnings for self-employed workers in NLSY 79 is as good as those in SIPP that were used in Hamilton (2000). Although Hamilton (2900) also used the earnings that include capital gain as an alternative earnings measure of self-employed workers. I just focus on the labor earnings of self-employed worker here. 9Hamilton (2000) evaluated the differences at 10 years of tenure and ‘20 years of experience. However. in the sample used in this analysis. only 1.54% of total sample has more than 20 years ofjob experience. Thus the difference was instead evaluated at 10 years of experience. About half of the observations in the total sample have more than 10 years of experience and about. 10 “/c of total sample has more than 10 years of job tenure. 23 of self-employed workers are higher without job experience or tenure as we can see from the positive coefficients for the self-employment dummy. However. the earnings- experience/ tenure profiles are flatter for self-employed workers than for salary/ wage workers. Because of the flatter earnings profile. self-employed workers with 10 years of job experience and 10 years of business tenure earn 111% (t = —1.731) to 23% (t = 1.709) less than salary / wage workers. depending on their educational background and marital status according to the OLS result. This result almost corresponds to the 19% self—employment penalty evaluated at 10 years of job (business) tenure and 20 years of experience found by Hamilton (2000) using labor income as self-employed workers” earnings. According to the fixed effects result. however. self-employed workers with 10 years of job experience and 10 years of business tenure earn almost the same amount as salary / wage workers on average. This result may seem to imply an upward bias of the OLS estimator: however. careful examination of the estimated coefficients reveals that the difference between the OLS and fixed effects results mainly come from the difference in the estimated coefficient for the interaction of self-e111})loyment and education. The first (‘lifference estimates of the return to education. however. are not reliable because the earnings before attent‘ling school may not reveal full individual lieterti)geneity (see discussion in Card (2000)). In particular. self-employed workers can attend school while they work less. which results in lower earnings. and then then start to work at their full capacity after graduating. which ultimately results in higher earnings. Thus the fixed effects estimate for the return to education among self—enmloyed workers may not be a reliable estimate and the conclusion instead should rely on OLS estimates. To summarize the findings. self-emrfloyed workers earn less mainly due to a. lower return to job experience and job tenure. When the difference is evaluated at 10 years of job experience and 10 years of job (business) tenure. self-employed workers earn 11% to 23% less of comparable salary / wage. workers on average. Next I attempt to explain this earnings difference with the conmensating earnings differential theory. 21 2.4 Job Satisfaction Scores and Descriptive Analysis The main survey item used in this study is global job satisfaction. The question reads How [do/did] you feel about. your job with [name of employer]? [Do/ Did] you(1) like it very much. (2) like it fairly well. (3) dislike it somewhat. or (1) dislike it very much'.’ (CODE ONE ONLY.) The distribution of responses for this question is tabulated in Table 2- 3. The distribution remains nearly constant over time. Examination of Table 2- 3 reveals that bout 65% of self-employed workers chose. "like it very much" while only about 45% of salary / wage workers chose this answer. It is also noTable 2- that only about. 35% of self-employed workers chose. “dislike,” ("dislike somewhat” and “dislike very much” combined.) while around 8-10‘fc of salary / wage workers chose this answer. This rough examination of the distribution clearly indicates that self-employed workers are more satisfied with their jobs. Before developing a detailed discussion of the compensating earnings differential based on job satisfaction scores. I examine whether this score is a meaningful economic variable. If the job satisfaction score contains meaningful information about. a worker’s actual job satisfaction. the score should predict the observed worker’s behavior. in par- ticular. the worker's future job change. Table 2- 4 tabulates the probability of job change between time t and t— 1 classified by the job satisfaction at time t — 1. This Table 2- clearly tells us that the worker who dislikes his/ her job is more likely to change his/ her job. Since the probability of job change may depend on the worker’s demographic. char- acteristics that may be correlated with job satisfaction. a probit model in which the probability of job change depends on demographic variables as well as job satisfaction is estimated. The results of the probit estimation appear in Table 2 - 5. Column 1. The re- sults indicate that those who dislike their jobs very much are 22% more likely to change their jobs than those who like their jobs very much. The inclusion of control variables hardly changed the results obtained in Table 2- ‘1. I have also tried the specification that included lagged log wage as an independent. variable, as used in Clark (2001). The estimated results for this specification appear in Table 2- 5. Column 2. The results indicate that a 10% increase in hourly earnings decreases the probability of job change by 0.26%. The size of this effect is relatively small compared with the effect ofjob satis- faction on job change. For example. changing job satisfaction from “like somewhat" to “like very much” decreases the job change probability by 10.8% (=0.231-0.l23). This larger effect. of job satisfaction on job change than wage effect can be explained as fol- lows. If the job match quality is more significant in job satisfaction determination than in wage determination. job satisfaction is a more crucial determinant ofjob change than wage because workers have a greater chance to improve their job satisfaction but. less chance to improve their wage through job change. This relatively small effect of hourly earnings on job change compared with job satisfaction was also found in Clark (2001), although the result in his study was not this extreme. \Ve should also notice that the size of the coefficients for job satisfactitm dummies hardly changed due to inclusion of log of hourly earnings. The results obtained from Table 2- ‘1 and Table 2- 5 clearly indicate that. the job satisfaction score contains valuable information about the worker's actual satisfaction in his or her job. One of the main drawlmcks of using the job satisfziiction score. as noted in the previous literature. is the difficulty in the interpersonal comparison of subjective measures. This study attempts to overcome this difficulty by using panel data because panel data t enables the researcher to examine the change in job satisfactitm associated with job change. In the follmving analysis. the job satisfaction score is assumed to be comparable within each individual over time. This assumption is much weaker than the assumption of inttiu'iwrsonal mmparability of subjective measures. As a simple way to examine the change in job satisfaction associated with job change, the transition matrices of job satisfaction for job stayers and job changers appear in Table 2- 6. Findings from these matrices are. smmnarized as follows: 0 Among job stayers. self-employed workers are more likely to stay in the "like very much" category as compared with wage / salary workers. (Panel A and Panel B) 0 Job changers who move from salary / wage jobs toself-en'iployed jobs are more likely to experience a positive transition of job satisfaction than job changers who stay in salary / wage jobs. It is also noTable 2— that job changers’ job satisfaction as salary/wage workers are originally slightly higher than stayers. This may be evidence of self-selection. (Panel A and Panel B) o This positive transition of job satisfaction is less likely to occur among those who change from self—enmloyed jobs to wage/ salary jobs. .\Ioreover, about 20% of job changers experience a negative transition of job satisfaction. (Panel F) These findings suggest the (:(‘inclusion that. self-employed jobs are more satisfying than salary / wage jobs. Hmvever. other demographic characteristics that also might af- fect job satisfaction. such as marital status. may vary at the time of job change and this might result in the findings above. To address this possibility, the effects of workers" observed and unobserved characteristics on job satisfaction are controlled in the follow- ing analysis. In addition. I attempt to calculate the monetary value of self-employment status in terms of job satisfaction. 2.5 Compensating Differentials among Self-Employed Workers To test the compensating differential hypothesis among self-employed workers. this sec- tion attempts to calculate the monetary value of self-employment status in terms of 27 job satisfaction to see if it is large enough to explain the earnings differential between salary / wage workers and self-employed workers. A straightforward way to see whether the difference in job satisfaction between self- employed job and salary / wage job explains the difference in earnings is to estimate the following equation: In 11'” = .30 + .1",-,.31 + .i)’-_>sclfa + .13, + :33jsl-t + cf, . (2.2) and see if H0 : .32 = 0 hold. If the null hypothesis is not. rejected. then we can conclude that the difference in earnings originated from job satisfaction. not from self-employment status. H(;)wever this obvious method neglects the fact that job satisfaction is also the function of earnings. as shown in previous studies. This endogeneity biases the estimates of .33 upward because j.s,~, and c” are positively correlated. This bias causes a. downward bias in the estimates of (.33 since job satisfaction and self-employn'iei1t status are positively correlated. as we saw in the previous section. This downward bias may result in the false acceptance of the null hypothesis. Thus. this avenue is not pursued in this paper. Instead, I examine how self-e111plt‘)yment status and earnings affect job satisfaction to see whether the countiensating earnings (’lifferential explains lower earnings among self- employed workers. To calculate the monetary value of self-e1nployment status in terms of job satisfaction, the marginal rate of substitution between self-e1nployment status and monetary earnings in terms of job satisfaction is calculated. The link between the job satisfaction score and utility is specified as iijI. 2 #3. if [13 > jsf, 2 112, if [.12 > jsf, 2111, if [11 > jsft. (2.3) JSit : idioms—x where jsi, is a catt-rgorical variable indicating worker 1'. at time is response to the job satisfaction question (1: "Dislike Very Much” - 4:"Like Very Much“ ), whereas js’f, is the latent continuous variable of job satisfaction and “(It 2 1. 2. 3) are the thresholds of job satisfaction that determine the answer for the job satisfaction question. Although many 28 factors may affect a workers job satisfaction. to see the tradeoff between self-employment. status and monetary compensation. those two factors are mainly considered as the determinants of job satisfaction. To estimate the monetary value of self—e1nployment. status in terms of job satisfaction, three additional assumptions are imposed. First. as a shape of job satisfaction function. a linear function is assumed as a first order a1.)proximation. Several (.lemographic variables are also assumed to affect job sat— isfaction. Moreover. other factors that may affect job satisfaction are assumed to be independent of self-emplt)yment status. monetary conmensation. and demographic vari— ables. and thus these factors are assumed to be normally distributed. This assumption results in Assumption 1 (Linear job satisfaction function) I“; = ()0 + 915;: + 92111 u‘n + rags + (3‘ + f’zt- (“ulna “7m J'ir- ('i N i ,(0‘ 1) i (2'4) where 8a is the dummy variable for self-e1nployment status. in, is hourly rate of pay. .1',-, is the vector of a workers attributes. and c,- is individual heterogeneity in utility level. Specifically .1‘” contains a. marital status dummy; a sex dummy: age. racial or ethnic group dunnnies: educational backgrmmd: labor market experience: and job (business) tenure. The next step is to calibrate the monetary value of self-employment status in terms of job satisfz-iction. As a. measure of monetary value, we can calculate how much workers can give up in terms of salary / wage earnings in exchange of one dollar earnings as a self- enmloyed worker while keeping their job satisfaction constant. The 0 in the following equation gives this ratio of trade off: 02111(c111',,)+.1',-,63+ c,: + (y, = gl + 92111 ”wit + (13,63 + c.- + 6);. (‘2.5) v js' of a salary/wage worker )5” Of a SGlf-GIIIployfi‘d WOI‘liCI’ When a worker receives 0 dollars of earnings per hour as a salary / wage worker, the worker has the same. level of job satisfaction when the worker earns one dollar per hour as a self-employed worker. The solution for the equation is simply, a = exp(01/62). In ‘29 other words. 0 dollars of earnings as a salary / wage worker is equivalent to one dollar of earnings as a self-employed worker. This value, a, is reported as a. monetary value of self-employment. status in terms of job satisfaction. To simplify the econometric i’nodel, two additional assumptions that will be relaxed later are made: Assimzptton 2 (Independence of Heterogeneity) c,~ _L 3;. try. .‘1,',- . (2.6) where s,- = [5,-1.51-2. ...,.s,-T]. w,- = ['1L',’1,U‘,‘2.....lU,‘T] and .L‘i = [.17i1,.l'7,'2,...,;1f,~7‘]. This as- sumption assures that individual heterogeneity is independent of observables and that the heterogeneity does not cause any inconsistency of the pooled, ordered probit esti- mator. Assmnptz’o'n. 3 (No feedback from current job satisfaction shock to future self-enmloyment status) 5,. U',‘..’17i.C,' ~1V(0,1) . (2.7) f’it This assumption rules out the feedback from current shock on job satisfactimi to future self—en'iployment status through job change. since if the feedback exists. the distribution of current 6 depends on future .9. These three assunmtions result in the pooled. ordered probit model. and the pa- rameters in (2.4) can be estimated. To estimate the model, I dropped the observations whose hourly rate of pay were either above the 99 percentile or below the 1 percentile in each year, and this sample selection results in sample (5) in Table 2- 1. Since those who earned extremely high wages and were extremely satisfied with their jobs could only report. “like their job very much" at. the maximum and vice versa for low wage earners. including those extreme earners would attenuate the coefficient. on hourly rate of pay toward zero and this would make the estimates of the monetary value of self—employn'ient status upwardly biased.10 The results of estimation appear in Table 2- 7, Column 1 and “’1 tried several trimming rules. The results were not essentially changed when I applied 5‘7t-95‘7t or 30 Column 2. The result. that appears in Column 1 is the specification that only includes the self-employment. diunmy and the log of hourly wage as independent variables. The coefficient for self—employment is 0.486. The size of the coefficient is not large enough to change the worker's response to the job satisfaction question from ”dislike very much” to ”like somewhat" or from "dislike somewhat.” to ”like very much,” since the critical values of the ordered probit are -0.377. 0.360. and 1.891. However, the size of the coeffi- cient is much larger than the coefficient. for log earnings. The coefficient for log earnings is 0.248. which is surprisingly small if we compare this value with the critical values. Due to this small effect of earnings on job satisfaction. one-dollar earnings of self-employed workers are evaluated as more than seven dollars of salary / wage workers’ earnings. This value is large enough to compensate for the lower earnings among self-en‘iployed workers whose earnings are about 20% lower than salary / wage workers‘ when workers have 10 years of job experimrce and 10 years tenure. When marital status. educational attainment. and job experience are included in the specification. the (.‘oefficient for earnings becomes even smaller because a part of high earnings is explained by the added explanatory variables. In this specification. the estimated monetary value of self—enn.)loyment status l)€('()ll’l(‘S a = 818; one dollar earnings as a self-employed worker is ermivalent to 88.18 earnings as a salary/wage worker. Again. this puzzling result is obtained due to the small effect of earnings on job satisfaction while the effect of self-employment status is large. 2.6 Extensions A surprisingly large estimate of the monetary value of self-employment. was obtained in the previous section. Now. I consider several possible reasons why the effect of self- employment status on job satisfacticm may be overestimated. To do so, I will relax the assumptions made so far one by one in this section. As partly suggested in the anal- ysis of transition matrices of job satisfaction, workers who become self-employed seem “IVs—90% rules. :3 1 to have a positive attitude toward their jobs independent from self-e1nployment status. If this is the case. the coefficient. for self—employed workers was overestimated in the pooled probit model. since c,- and s" are positively correlated. In addition. if workers with high ability have high expectations for their earnings, workers with high ability are less happy with their current earnings. As a result. unobserved heterogeneity and current earnings. which is a. proxy for ability. may be negatively correlated and the co- efficient for hourly rate. of pay may be negatively biased. Considering this heterogeneity or other possibilities that unobserved heterogeneity in job satisfaction are dependent upon observable characteristics. the assumption 2 (Interpersonal comparability of job satisfaction) is replaced with Assumption 2' (the "Fixed Effects" assumption) - __ — _ -) ‘ . (ylsi. ti’,..1',- ~ .\ (ms,- + ”,gln (1',- + 73.13. 0;). (2.8) This assumption allows depemlence between c,- and 3,: or .r,- in a restrictive way.11 where §,-. lnfuy and f,- are mean of s,-. In 11',- and 1',- respe(’-tively.12 The consistent es- timators are obtained through a. pooled. ordered probit estimation of the model that. includes individual means of lll(l("})(’ll(l(‘llt variables. The importance of assumption 3 should be emphasized here. If current. shock to job satisfaction. e”. affects the future value of self-em])loyment status through job change behavior, e” and .57.- , which is one of the independent \I’ariables and dependent on the future self-employn'ierit status. is dependent and a consistent estimator cannot be obtained. The results of estimation appear in Table '2- 6. Column 3. The coefficient for self- ernployment decreases as expected from the positive correlation of c,- and 5,1,.At. the same time, the coefficient. for earnings increased due to negative correlation of ed and In 1.1.3,. As a. result, the calculated monetary value of self-einplr.)ymeiit status becomes 284% of earnings. llMundlak (1078) proposed a variant of this assumption in a linear regression framework. 12The analysis under this assumption is called a fixed effect analysis because this assumption allows dependent unobserved heterogeneity.. As confirmed in the previous section and in previous studies (Freeman (1978). Clark. Georgellis. and Sanfey (1998) and Clark (2001)). job changes tend to follow low job satisfaction. In the light of this fact. ruling out the feedback from e.“ to 55,4. u-Sn‘ is a strong assumption. In particular, if a salary/ wage worker experiences low job satisfaction because of some shock (after conditioning on individual heterogeneity) and become a self—employed worker. we tend to overestimate the effect of self-employment. on job satisfaction because a salary / wage workers with current negative shock is more likely to be self-employed in the following period. To take care of this possibility. feedback effects are allowed through the assumption. Assumption. 8° (Feedback effect) ('-,:,|s.-. .1?,:.(1,- ~ .\'(h(s,-,.s.~,r1. H.817). 1). (2.9) where h(s,~,. sits]. 8,7”) = (5111121.X(l8,-,*T — .9,,+T_1|. lsit+1 — .s,,|). This function in- dicates that if a change in self-employment status takes place within T years due to current shock on job satisfaction. the parameter 7' is the maximum lag period of the feedback from c” on future job change. For example. if 7' = 3 then the feedback is assumed to take place within 3 years. The cases 7' = 1. 2 and 3 are considered in the following analysis. The model is estimated with a pooled. ordered probit model with an individual mean of independent variables and h,(.) as independent variables. The estimated results of the "fixed effects" probit model with feedback effects appear in Table 2- 7. Columns (4) - (G). The estimated coefficients for the feedback term. h(.). are all negative but. they are not significant. in all specifications. Negative coefficients im- ply that a current negative shock on job satisfaction causes a change in self-ernployment status through a job change. However. even after considering correlated heterogeneity and feedback from job satisfacticni to future job changes. we find essentially similar re- sults as before: The estimated monetary value of self—er’nliiloyinent status is about 390%. Since the coefficients for the feedback terms are not significantly different from zero, I take the fixed effects estimate. which is d : 3.387(s.e. = 0.097). as the most preferable 33 estimate for the monetary value of self-e1nployment status in terms of job satisfaction. Now. the very high valuation of self-employn'ient status is the puzzle that should be explained. Although self-employed workers with 10 years of job experience and 10 years of business tenure earn about. 20% less than their salaried / wage—earning counter- parts. the estimated results imply that one dollar of earnings as a self-employed worker is equivalent. to three to four dollars earned as a salary/ wage worker in terms of job satisfaction.13 If we consider job satisfaction as equivalent to utility. this value means that self-ernployed workers do not move to salary / wage jobs even when they are offered three or four times more than their current earnings. but. this finding is counterintuitive. However. there are four explanations that may explain these surprising findings. The first. is the fact that job satisfaction is only a segment of utility function. Sup- pose the simplest form of utility function. which consists of only consumption and job satisfaction: -u(c(zc).js(se. w)) . (2.10) where c is consumption that is presumably a function of earnings. Then the marginal rate of substitution between self-e111})loyment status and earnings is UH ~ . . (l '. ° 18.. 4 V _ ()M‘ _ J5 - M U“. (1.. - (u. + U}. ~Js... where f_,. denotes the partial derivative of f with respect to .1'. Since (1.. > 0 and cu. > 0. the monetary value of self—e1nploymmit in terms ofjob satisfaction. which is (j.5'..,.(../j8u.). overstates the value of self-employment status in terms of utility. Although the mone- tary value of self-em})loyment status in terms of job satisfaction has been estimated in this paper. the monetary value of self-e1nployment status in terms of utility should be estimated to appraise the validity of the (“:ompensating differential hypothesis. However, calculating the monetary value of self—ernployment status in terms ofjob satisfaction is a. useful exercise in light of the reality that a numerical measure of utility is not available. 13Althorugh I tried several specifications in which the monetary value depended on job experience and tenure. the result virtually did not change. The second explanation concerns the overestimation of success among self-employed workers. Empirical studies have found that self-employed workers are overly confident in their future success as compared with salary / wage workers (Cooper. “'00. and Dunkel- berg (1988) and Arabsheibani. de.\Ie/.a.. Maloney. and Pearson (2000)). For example. Arabsheibani. deMeza. Maloney. and Pearson (2000) report. that self-employed work- ers expect better financial outcomes in the following year than salary/ wage do, even though they. in fact. experience worse outcomes. \\'hen self-employed workers expect. future monetary success. self—employmei1t status has a subjective "option” value for future earnings and this option value may make self-enmloyed workers more satisfied with their jobs. Although this effect is attributed to the non—monetary value of self- employment. status in this study. it instead should be attributed to the mcmetary value of self-e111})loymeiit status. This overestimation of the non-monetary aspect and un- derestimation of the monetary aspect result in an overestimation of the non-monetary aspect of self—emploved jobs in terms of job satisfaction. The third explanation relates to underreportii1g in self-elliployed workers’ earnings; if this is the case. the value ofself-e111})loyment is overestimated. since the utility from underreported earnings is captured through self—employment status. Although NLSY makes good efforts to collect reliable labor income from self-employed workers. self- employed workers may refer to their tax forms from previous year in which earnings were uiiderreported for tax avoidance pin‘poses when they answer earnings questions. For example. Joulfaian and Rider (1998) show that the underreport rate of self-employed earnings is about 20% on average. using the Taxpayer Compliance Measurement Pro- granrl'lAlso. self—employed workers may consume out of business expenses. For example. they may drive coi‘npairv—owned cars for personal purposes. This also may increase the monetary value of self-employment status. but it. should not be interpreted as a com- pensating differential since it. simply captures consumption. HThe Taxpayer Compliance Measurement Program data are stratified samples of individual tax returns subject to intensive line-by—line exz—uninaticms (.loulfaian and Rider (1998)). The fourth explanation relates to i heterogeneity in the marginal job satisfaction from self—employment.. “'hen heterogeneity is explicitly modeled, the utility function becomes J“; = 90 + (9115i! + 92111 Ur'u + 1,193 + Ci + Cit- Z 90 ‘i’ 918,7 + 9211111”); + J'it93 ‘i‘ Ci + (’1'), THIS”. (1')). .I'); N .‘\Y((91,'— 01)5i(, (TC). (2.12) hvlai‘ginal job satisfaction and self-em;)loyment 01,-, and .s,-,, are likely to have a positive correlation since those who gets higher job satisfaction from self-employn'ient are more likely to be self—employed. consequently. s” and the error term is positively correlated and plime—li > 91. Thus the pooled. ordered probit estimator overestimates the average of 61. This positive correlation can be very large for workers who want to be self- employed at any cost. Thus the calculated monetary value of self—employn’ient from the pooled probit model can be interpreted as the upper bound of the average evahiation of self-employmeiit. throughout the population. In addition. the entry to self—employ1nent continues so long as a workers evahiation of self—employment. status is above the earnings penalty of self-employed workers. Thus. the earnings penalty of self-employed workers is determined at the margin. through a market mechanism. It is no surprise to find a higher average evaluation of self-employment status than the evaluation of self-employment status by the last, worker who marginally becon‘ies a self-employed worker. The interpretation of the monetary value of self-employment status becomes more re- stricted for the fixed effects model. The coefficient 0“ is identified as those who changed self-employment status during the sample period since if Sit is constant over sample pe- riod, the variable is perfectly multicolinear with .57,- and only 01i+71 is identified for those observations. Thus the estimated Incmetary value of self-eniployment status from the fixed effects model is the average (waluation of self~e1nployment status among workers who experience a transition between a salary / wage job and self—employed job. As we can see from Table ‘2- 6. there are more observations that transit from a salary / wage job to 36 a self-employed job. so it is not surprising to find higher evaluations of self-e1nployment status among workers who become self-employed during the sample period. Those who leave self-employmeiit might also do so because of financial reasons but not because. of job satisfaction. These four considerations may well reconcile the estimated self-eniployment penalty (20% of earnings for a worker with 10 years of job experience and 10 years ofjob tenure) with the estimated monetary value of self-employinent jobs (300% - 100%) in term of job satisfaction. 2.7 Conclusion Analysis of job satisfaction scores show that. self-employed workers are more satisfied with their jobs than salary / wage workers. .‘\Ioreo\-'er. one dollar of earnings while a self- employed worker is equivalent to three to four dollars of earnings while a. salary/wage worker in terms of job satisfaction. This finding is preserved even when individual het- erogeneity. which is potentially correlated with self-employinent status, and the feedback effect. which runs from current job satisfaction to future job change. are considered. This high valuation of self-emph)yment status in terms of job satisfaction may over- estimate the actual trade-off between self-employment status and monetary income in terms of utility. However. even after taking the effect of unavoidable overestimation into consideration, the value of self-em})loyment status in terms of job satisfaction, which is 300 to 400‘} of other workers' earnings. seems high enough to explain the lower earnings of self-employed workers. Thus. the results obtained in this paper support. the compen— sating differential hvpothesis as an explanation for lower earnings among self-employed workers. Promising future research would develop a rigorous appraisal of the con'ipensating wage differential hypothesis by using a better measurement of utility or principals of revealed preference. 37 Chapter 3 Peer effects on substance use among American teenagers 3. 1 Introduction Widesmead use of illicit substances by American teenagers attracts both public attention and research interest. The changing percentages of substance users during the 1990s are plotted in Figure 2-1. Although the percentage of alcohol users dropped in the early 1990s, it still remains high. This figure shows the steady trend of cigarette users at a high level.1 It is also notable that the percentage of marijuana users increased from 4.41% to 9.7% l)(~?t\~\'t‘(‘.ll 1990 and 1997. These figures have sparked much public interest about the reasons why teenagers use substances and what policy makers can do to reduce this usage. Besides the price of substances, peer effects or peer pressure is identified as a. critical (ilet.er1ninant., since the use of substances is considered to be a. highly social bel‘1a.Vi(;)1‘.“2 Reacting to this interest, economists and sociologists have tried to estimate the ex- istence and the strength of peer effects. Idei‘itifying peer effects is not. easy since an observed behavior shared liiiy a teenager and his/ her peer may result from unobserved factors that group members share instead of peer effects. In addition, identifying peer 1On the other hand, Gruber (2000) reports that the percentage of cigarette smokers increased by one-third between 1991 and 1997. (Based on youth behavior risk survey. for 9th through 12th graders. the number has increased from 27% to 36% between 1991 and 1997.) 28ee Los-Angeles—Times (1999) for interviews of youth smokers on the reasons why they smoke. 38 effects becomes complex when the average reference groups outcome is used as a mea- surement. of peers’ behavior. Determining whether a teenager's behavior affects his/ her peers or vice versa is difficult. Manski (1993) articulated this as the ” reflection prob- lem.” In addition, both current substance users and backgrounds of group members may affect individual behaviors. Although both effects are called peer effects, each has dif- ferent policy implications. Distinguishing between these two effects, however, is known to be difficult (Manski ((1993)). This paper ermiloys a critically (.lifferent strategy for identifying peer effects; I identify peer effects by using teenagers’ subjective perceptions. Manski (1993) wrote: Given that identificatimi based on observed behavior alone is so tenuous, ex1.)erin'1ental and subjective data will have to play an important role in future efforts to learn about. social effects. Using subjective pt-u‘ceptions of peer behaviors, identifying peer effects is free from the problems that arise with using an average outcome as a peer variable. According to Manski (2000), this approach has not been taken seriously since he originally suggested it in 1993. In addition, I employ school and household fixed effect estimation to ensure the rolmstness of my results. 3.2 Reflection problem Manski (1993) articulated several issues concerning the it'lentification of peer effects using an average group outcome as a peer variable. A short sketch of the problem follows. Let y be an outcome of interest. I be attributes characterizing an individual’s reference group and 3 be attrilmtes that affect the outcome directly. The outcome is characterized by y = a + 3E(y(.r) + E(:|.r)7 + :1) + u. E(u|.r, :) = .175. (3.1) 39 If :3 7E 0. then an individual behavior is affected by the mean of the group outcome, E (ylr). This is the ”endogenous effect.” If 7' 75 0. an individual behavior is affected by the group mean of the exogenous variable (background of group members); this is the "' contextual effect.” If 6 ¢ 0, the model exhibits the ”correlated effect.” People in the same group behave similarly because their shared group characteristics are correlated with unobservable factors. such as social institutions. Taking conditional expectation on 2 and 1‘, the model becomes E(y|.r. :) 2 (1+ .3E(yl.r) + E(z|.1")",r + 2!) + 16. (3.2) To discuss the identification of parameters, we need to solve for the conditional expec- tation of y in terms of .r and 2. Using the iterated law of expectations, E(y|.r) = (r + .3E(y|.r) + E(:|.1‘)‘," + E(Z|.1‘)'I] + 1‘6. (3.3) Solving the expression for E ( ylr) and substituting it into (3.2), we obtain E(y|.r. 2) = (,1/(1—3)+ E(:|.l')l(7+;31])/(1—— 3)] +.r6/(1—,3)+ :n. (3.4) The composite parameters are identified if [1, E (2)17), 1?. z] are linearly independemt.3 Even if the linear independence is assured, the endogenous effect ( 113) is not identified if the contextual effect. is present. (La 7 7f 0). As Manski (1993) stressed, distinguishing the endogenous effect from the contextual effect is important. because these two effects have critically different policy implications. Consider, for example, that a lecture on the health effects of smoking is provided to one class but not to the other classes in a particular school. If the lecture effectively makes students quit smoking in the class. 3The linear independence of explanatory variables is violated if 1. : is a function of .1: since E(:(.r)i.r) = Z(.‘l'). Consequently E(:|.r) and z are linearly dependent. This case occurs in the situation when group members share the same exogenous variables. 2. 3 does not vary with .17. Since E(:).r) is constant. E( :|;r) and z are linearly dependent. This case occurs if average exogenmls variables are identical across groups. 3. E(:l.r) is a linear function of .1: since E(:).r) = 0.1:. This is a theoretical possibility rather than an actual possibility since .17 is usually an index for a group. ‘10 the effect. propagates to the students in the other classes through the endogenous effect. If the endogenous effect does not exist, the effect of the lecture is limited to the class where the lecture is given. \Vhile the endogenous effect implies this ”social multiplier,” the contextual effect and the correlated effect do not imply that. Moreover, .\Ianski (1993) warned that, in general. sample correspondence of E (yl.r), denoted by E,\-(;y|.r). is not identical to E ( glr). As a result, [3 can be calculated in the sample, even if 3 is not identified in the population. Thus, the successful calculation of 13 does not. imply anything about the identification. After all, the identification of the mixture of the endogenous effect and the contextual effect is possible under the assumption of the linear independence of [1, E (: .r).;1:. z]. Distinguishing between the endogenous effect. and the ccmtextual effect is. l’lUWE’VCI‘. principally impossible. The difficulty of the identification arises because the mean of the outcome is used as an explanatory variable. I avoid this complication in this study since the direct subjective perceptions of peers’ behavior, instead of the average behavior of peers, is the key explanatory variable. The beauty of this approach is that, once the linear independence assumption of explanatory variables is assured. the endogenous effect and the contextual effect. are Scip(l.’l'(lf(:fly identified. In the next section, the model with subjective perception is discussed. 3.3 The model with perceived peer behaviors The model with perceived peer behaviors is specified as follows: y 2 (1+ :3p+ E(: 1))" + :57) + 21., (3.5) where y is an outcome of interest, .r is attributes characterizing an individuals reference group and 2: is attributes that. affect the outcome directly. The variable [i is subjec- tive perception of peer behaxv'iors. Once E ( ulp. 2,17) = 0 and linear independence of [1. p. z, E (2);1:)] are assured,4 the parameters, a. .7 and I}, are identified through OLS. ‘JThe assumption of linear indepemlence here is more restrictive than the assumption needed in Manski (1993). since 1) is newly added to the list of \I'ariables which should be linearly independent. 51 1 Since the group average of observed behaviors is not used as an explanatory variable as in Manski (1993), no complication of identification arises. The crucial assumptitm in this model is that perceived rather than actual peer l.)eliav- iors determine individual behaviors. This assumption is consistent with the conformist model. Akerlof (1997) argued that individuals get, utility from behaving like an ”aver- age" person in a reference group. It is natural to assume that perceived peer behaviors produce the image of the ”average" person. This ”conformist” behavior may be rein- forced through the formation of group norms. By forming group norms that require members to engage in similar behaviors, each individual group member enjoys higher utility. Becker (199(5) considered a utility function (U) that has own norm (N) and norm of peers (Np) as arguments and its cross derivative to be positive (62b'/(é3i’\701\’p) > 0). Thus individuals can obtain higher utility through forming their own norms similar to the group norm. It is natural again to assume that each individual perceives group norms through perceived peer behaviors. Individuals bel'iave according to their norms. To summarize, the assumption that. perceived peer behaviors detern'iine individual be- haviors is a natural one if peer effects operate through ”conformist” weft—arences or the enforcement. of group norms. At the same time. there are many other reasons why peer effects exist. For example, American kids play baseball instead of cricket since other kids know how to play baseball, and, consequently, it is easier to find playmates. .\~loreover, since playing an actual game would be much more fun than playing catch. it may be an important goal for kids to find many playmates easily. In this example, utility obtained from playing baseball depends on the number of players with whom a kid plays (up to 18 players). By the same token, teenagers may obtain higher utility from tobacco if they consume it with their friends. In these situations. actual peer behaviors. rather than perceived ones, affect an individual’s behaviors. Although this possibility cannot be ruled out, it is simply assumed that, an individual behavior is influenced only by the perceived peer behaviors.5 Relaxing this "Of course. the actual peer behaviors may determine perceivetl peer behaviors: ll()\\'(‘\'(,‘l‘. here I 42 assumptitm causes complication in identifying the endogenous effect as noted earlier. Thus, the interpretation of the endogenous effect in this paper should be restricted by this assumption. 3.4 Literature review Many papers shed light on peer effects. In this section, the papers are classified by the identification strategy used. The identification of the endogenous effect is especially focused because of its unique policy implication of the ”social multiplier.” A thorough review of the existing literature reveals the strengths and weaknesses of each identifica- tion strategy. In particular. the limitation of an identification strategy that only uses observed behavior becmnes clear.6 I also introduce the studies that use economic theory to help provide additional insights into the identification of the endogenous effect. It should be noticed that each study does not necessarily fall into a single category since several identification strategies may be used in any given paper. 3.4.1 Identification through proxy variable Several studies use variables that represent "peer quality” or ”neighborhood quality.” The problem of whether the endogenous effect can be identified using these I)I‘(_)Xy vari- ables is considered. The outcmue of interest (teenage pregnancy and high school drop- ping out in Evans, Oates. and Schwab (1992) and high school dropping out in Crane (1991)) is cliaracterized by y = a + .3E(yl.r) + 211+ '11. (3.6) assume that once perceived peer behaviors are included in the behavioral equation. the equation does not include actual peer behavior. 6Many of studies surveyed in this section used nonlinear models such as probit. However. we discuss identification strategies in the context of linear models, since the discussions of identificaticm fundamentally carry over. Moreover. the identification of parameters should not depend only on the ncmlinearity assumption. The sizes of peer effects in those studies are discussed in terms of marginal effects. "13 Since two of studies above implicitly assume absence of the contextual effect, I also adopt that assumption. Instead of E (ylr). however, a proxy variable, such as the portion of students who is eligible for a. school free lunch program (Evans, Oates. and Schwab (1992)) or a neighbm‘hood occupational structure (Crane (1991)), ([2 E(yl.r) + 't‘ (3.7) is used in the estimation. The equation actually estimated reduces to y = o +.3(1+ :I; + u — .31‘ (3.8) The assumptions E(u|.r. :) = 0 and E(c|.r. :) = 0 are required for the identification of the endogenous effect. ,3. Assuming the absence of the correlated effect, the first assumption is assured. ()n the other hand. the second assumption is questionable. Since q in these studies consists of variables defined on reference groups, q are naturally functions of .r. As implied by (3.7), '1' is a function of .r accordingly. Thus, E (elm. z) 7é 0. Although using proxy of E (glr) reveals something about broadly defined peer effects, identifying the endogenous effect is impossible using a proxy variable. 3.4.2 Identification assuming the absence of the contextual ef- fect Many studies of peer effects fall into this category. Case and Katz (1991) looked at the effect. of neighborhood average incidence on youth behavior, including drug use. Sacer- dote (2000) looked at the effect of randomly assigned roommates in a college dormitory on student outcomes such as GPA. Gaviria and Raphael (2000)) looked at peer effects within schools on students' behavior.7 These studies try to identify the structural parameters in (3.2). As we have seen in (3.4), the endogenous effect. (.3) cannot be identified without assuming absence of the contextual effect (i.e. '7' = 0). Consequently, all of these studies assume the absence 7The behaviors considered in Gaviria and Raphael (2000) are substance uses, church attendance, and dropping out of high school 11 8 Replacing the population mean of the contextual effect except for Sacerdote (2000). E(y|.r) with the sample average E,\v(,z/|.r), the (3.2) is estimated assuming '7' = O. The assumption '7 = 0 is not severe if the research interest. lies in examining broadly defined peer effects. However. if the existence of the "social n'iultiplier” is the main concern, the assumption is restrictive. since all of the observed peer effects are attributed to the endogenous effect. by assumption. Thus the identification based on '7 = O is not appro— priate for deriving policy implications since the impact of policy intervention depends on the existence of the "social n‘iultiplier.” Most of the studies mentioned earlier were concerned about the violation of E (u Ir. 2.) = 0 and the consequent bias in the OLS estimator. Unobserved group characteristics (e.g. teachers” competence in a school) make E (ulr. z) a function of .17 (the correlated effect). This correlation of .r and it makes the OLS estimator of 3 biased since E ( ylr) is also a function of 1'. Alternatively, unobserved individual characteristics correlated with ob- served characteristics make E ( (1.).1'. 2) be. a function of z. This correlation makes the OLS estimator of I] biased and the bias could be transmitted to the estimator of 3. For the latter concern, IV estimation implied by (3.3) solves the problem. Under assumption ’7 = 0, E ( 2|.r) is excluded from the behavioral equation of interest. (3.2) but correlated with E(y|.r) given I] # 0; thus E(:l.r) serves as instruments for ELI/[1). An important point here is that u is allowed to be correlated with 2 but not with .1:. If u is correlated with .r, E ( zlr) is correlated with u and cannot be an IV. Evans, Oates, and Schwab (1992) considered the case in which 11 includes parents’ ”willingness" to invest in their children when this willingness affects where the family resides (at) so that the parents choose peer quality (E(y).r)) endogenously. In this sit- uation of endogenous sorting, u and .r are correlated. To address this concern, they 9 used varieties of E (2);!) as IVs. As criticized in sulmequent studies, this strategy is 25In Sacerdote (2000). additional information of error structure was used to identify :3 and “)- as we. will see in the next subsection. He found 5., to be negligible. After this finding, he used 7 = 0 to identify the endogenous effect. 9The metropolitan area unemployment rate. median family income, poverty rate. and the percentage of adults who completed college were used as 1V s. dubious because the correlation of u and .17 (caused by endogenous sorting) implies the correlation of u and E (:|.r). Thus E(:|.r) is not ideal IVs for E (ylr). Accordingly, their conclusion that peer effects do not exist. after controlling endogenous sorting is dubious as well. Sacerdote (2000) used randmn assignment of roommates in a college dorm to avoid this endogenous sorting. This natural experimental situation effectively rules out. en- dogenous sorting. Gaviria and Raphael (2000) divided the sample into two sub-samples: families that. had moved in previous two years and the families that had not moved. They argued that if endogenous sorting is widespread, estimates of peer effects for families that had recently moved into a new neighborhood should be larger. Their findings were mixed.10 Although this method does not capture endogenous sorting that had occurred more than two years earlier. the impm'tant issue here is whether peer effects are critically different across groups. This is a clever possible strategy when the random assignment of peers is not available. 3.4.3 Identification through variance covariance structure The identification strategies intrmluced so far all assume the absence of the contextual effect (7 = 0). To the best of my knowledge. the only paper that attempts to relax this assumption is Sacerdote (2000). Since there are only two students in a dorm room, the system of interest is g) 2 (1+ :3”)- + 3.1-c," + 3,!) + u,;, (39) yj = a + {3.1/1: + 3r," + zJ-I} + “j. (3.10) where Ire/((1,) = lv'u/(uj) 2 (TS is assumed and E(u,|:,, 31-. uj) = E(uJ~|:,-, 31.11,) = 0 is assured from the random assignment. Substituting (3.10) into (3.9), the reduced form (I + (1.3 (1+3)? (1+ 3)!) 11;+._311j ‘U’: 1—.3‘3 2 1—.3'-’ 31—32 1—32 (3.11) ‘ . . . o. . I . l’Different peer effects were found in the two sulrsamples for marijuana usage but not in drinking, smoking. church attendance or dropping out of high school. .16 is obtained. Estimating this equation through OLS. three coefficient estimates are pro- duced. Since there are five parameters (a. 3. 7, r). (73), two more estimates are required to recover all of them. The remaining two functions used for identifying the parameters come from the variance and covariance structure of error terms due to the randonmess of the roommate assignment. Then the variance and covariance structures are u-+.3u- 1+32 ., l' - ._'___-’. = —— , ‘ 3.12 m( 1_ J") ) (1_,j2)‘20“ ( ) ktlltl (‘or(lli + .311}. 111- + 311,-) = 2.305 . (3'13) 1— 32 1— .32 (1 - 3‘3)? Using three coefficients estimatm’s of the OLS and the estimates of variance and co- variance of residuals. parameters ((1. {37:11.03) are recovered.” Comparing the two estimates of .3 obtained through this method and (i)btained assuming e,- = 0,12 he found similarity between the two estimates and concluded that the, peer effect. works through the endogenous effect rather than the contextual effect (i.e., ’7' = 0 is not restrictive assumption.) Although this is a path breaking approach, the identification crucially depends on the homoscedasticity assumption embtxlied in (3.12) and (3.13). It is not clearly a priori, however, if all of the. students in dormitories share equal variances in the unobservable characteristics. As in the usual discussion, the identification off the assumption of error structure is rather tenuous. Results obtained in this study thus are informative but not definitive. It. should be noted, however, that this aptn'oach is the best. one can take with a data set that contains only observed behaviors. 3.4.4 Identification through dynamic structure Manski (1993) introduced some studies that exploit the dynamic structure of peer effects transmission. The estimated equation is the dynamic version of (3.2): E,(g|.r. :) = o + .3E,_1(g|.r) + E,_1(:|.r)7 + 2,1} + 1,2) (3.11) 11Standard errors of flame estimates are calculated through bootstrapping. 12Even under assumption of A, = t). regressicm of g, on y_, and s,- does not render unbiased estimator of (r. .3. 7/ and 0?, since Et’u,'g_,. 3,) ,i (1 since y, is a function of 31,11] and u, as shown in (3.11) even if e, = (1. His discussion is not quite correct in this sense. 17 Since peer effects do not operate conte1npm‘aneously, the estimation is free from the ”reflection problem" and opens the possibility for estimating peer effects. Studies of Biddle (1991) and Norton. Lindrooth, and Ennett (1998) fall into this category. Biddle (1991) analyzed peer effects in the demand for personalized license plates using state-level aggregate data. In his study. 3;, is current demand of personalized license plate in state .r. E,_1(y|.r. :) is the demand in the last year. 2, includes state aggregate income and the number of cars in a state. State dummies are included as .r,. He implicitly assumed "y = 0. He found significant peer effects in the demand. Norton. Lindrooth, and Ennett (1998) analyzed peer effects in Sixth to ninth graders“ substance uses. In their study. 3], is sixth to ninth graders’ current use of substances, and E,_.1(y|.17) is average use of substances among students who attended the same primary school. The vector z, included sets of demographic and regional characteristics. They imI')licitly assumed 7 = 0. Since peer effects are defined at an earlier schooling level. the estimation of peer effects is the dynamic version. The result of estimation showed that 3 is in the neighbor of 1. Although these estimators are immune from .\Ianski (1993)'s criticism, one should realize that identification of .‘3 crucially depends on the exclusion of contemporaneoi1s peer effect. by assumptions. Once the assumptions break down, Manski (1993)’s criticism again applies to these two studies. It is not clear whether the contemporaneous peer effect can be ruled out in those two studies a. priori. 3.4.5 Identification through sibling method Aaronson (1998) attempted to address the heterogeneity bias caused by household het- erogeneity such as parents‘ ”willingness” to invest in their children. He estimated the effect. of neighborhood quality, measured by poverty rate or average dropout rate. on a teenagers dropping out of high school. His model specifies outcome as my, 2 a + E(.’/l'1.)fj."3+ 3,1,1) + ('j + um. (3.15) ‘18 i,t and j index individual. time. and household. respectively. E ( yl'J'), J- is a regional av- erage and (:j denotes an llll()lf)S€I‘V0(l household heterogeneity. Absence of the contextual effect was assumed. The sibling method differences out cj but it also differences out E (y|.r),j if it is time invariant. Thus. the identification of the peer effect, 3, crucially depends on the time variance of E ( ylr)” along with the time invariance of household het- erogeneity. cj. To obtain variation in E ( ylr)”, he used a sibling sample from households that moved. As Aaronson (1998) noted. however. the residential change and change of family’s unobservable components are likely to be correlated,13 and consequently the time invariance assumption on cJ- may be violated. .‘VIoreover. the change of :17 should af- fect individual behavior only through the change of E (ylr) (i.e.. the assumptions ’7' = 0 and E("u|;r. z) = 0 are critical). Regardless of these restrictions, it is informative that he found significant peer effects even after controlling household heterogeneity, since the bias caused by the household heterogeneity was critical concerns in previous studies. The sibling method is also enmloyed in this study to identify the endogenous effect; however, time invariance of household heterogeneity is no longer needed by using sub- jective perceptions of peers' l_)ehaviors. since the siblings" perceptions may vary within a household. Accordingly. c J- can be differenced out without differencing out perceptions. 3.4.6 Identification through economic theory So far. I have surveyed identification strategies that only use observed behaviors as information. On the other hand, there are several notable studies that identify peer effects using prior kntm'ledge suggested by economic the(_)ry. Neumark and Postlewaite (1998) suggested ” relative income concern" as a factor to explain the rapid increase of the labor force participation (LF P) rate among US. married women. C(_)11stru('-ti11g an economic model of ” relative income concern," they actually estimated peer effects on labor force participation among married American women. They used a sister-in-law's LFP as the peers" li)ehavior and regressed a woman’s 13Negative unobservable shock within a family may make the family move. 19 LFP on her sister-in-laws LFP. They found significant peer effects on the LFP decision. Moreover, they directly estimated the prediction obtained from their theory. The theory predicts that. a woman is more likely to be employed if her husbands income is less than her sister-in-law‘s husbands income and the sister—in-law is not employed. This is because the woman’s household may "win" in the income race due to the woman's additional earnings. ()n the other side of coin, a woman is less likely to be employed if the woman with the lmsband whose income is less than the sister—in-law‘s and her sister-in-law is employed. This is because the household is unlikely to "beat” the sister- iii-laws household with the woman's additional earnings. The estimated result of this ” best response function" was consistent. with theoretical prediction. Munshi (2000) also avoids spurious findings by using predictions from economic theory. He analyzed the teclmological diffusion during the Green Revolution in India. He used the teclmological characteristics that the High Yield Variety (HYV) of rice is more sensitive to unobservable land quality than is the high yield variety of wheat. As a consequence of this teclmological factor, economic theory predicts stronger peer (social) effects for the diffusion of wheat. This is because, roughly speaking. an agent cannot learn much from his/ her neighbors experience when the success of technology adoption strongly depends on unobserved heterogeneous factors. The data supported this prediction; he found stronger peer effects in the wheat. HYV diffusion than in the rice HYV diffusion. These studies are persuasive because they carefully examined theoretical predictions and did more than regress the individual outcomes on group outcomes. Accordingly, these studies are free from .\Ianski (1993)'s criticism. This strategy. however. can only be applied to the situation in which sharp theoretical predictions are available. 3.5 Estimation This review of the existing research clearly indicates the limitation of studies that use only observed behavior to identify the endogenous effect. As has been discussed, eco- nomic theory can provide precious information for the identification of these effects. In the context of substance usage by teenagers. however. obtaining such a. theoretical prediction is difficult since the preferences of teenagers. which are necessary to derive behavioral predictions theoretically. are largely unknown a. priori. I thus use perceived peer behaviors. which are potentially error ridden. as additional informaticm to identify the endogenous effect. The description of data. the econometric model. and the results of estimations follow. 3.5.1 Data The data set used in this study is the National Longitudinal Survey Youth 97 CeoCode file. The sample construction is summarized in Table 3-1. I used the set (10) (N=6356) as a sample for the cross-sectional studies. and I applied (almost) school fixed effect es- timation and sibling fixed effect estimation to the set (11) (N26312) and (12) (N22158) respectively. In Table 3—1, the sample means of outcomes are tabulated. From the tabulation, we can confirm that the sample selection does not. drastically change the property of sample in terms of outcomes. The outcomes are constructed by using questions about substance use in last 30 days. The respondent who smokes / drinks more than or equal to one. cigarette/ drink is defined as a smoker/tlrinker; similarly. the respondent who uses marijuana more than or equal to once per month is defined as a marijuana user. In order to construct the peer variables. res1.)(_)ndents were asked about their subjec- tive perception of their peers' behavior by the f(.)llowing questions: ”\Vhat percentage of kids (in your grade / in your grade when you were last in school) (smoke / smoked ) cigaret tes'.’ (get/got) drunk at. least once a month? (have / ever)used marijuana. inhalants. or other drugs?" Respondents were allowed to answer the questions with one of the following five categories: 1. almost none (less than 10%) about 25% 9°!0 about half (50%) about 7.5% :L“ 5. almost all (more than 90% ). From these categories. I constructed perceived peers’ behaviors in which ”almost none” was coded as 0. and ”almost all" was coded as 1. Descriptive statistics of individual substance uses and perceived peer substance uses are tabulated in Table 3—2. An interesting finding is that the respondents systematically overestimate peer behaviors and the degree of overestimation is not negligible. It is worth noting that this HIPHSIII‘PIIIE‘IIT error is not a problem if we assume the variable that affects the respondents’ behavitn's is the perceived peer behaviors rather than the ”objective" (for econometricians) peer behaviors.” 3.5.2 The model The substance use by a teemigm' is specified as y = n + 3’1) + E(ZIJ‘)"; + :1} + u. (3.16) where y is a binary variable set to one if the individual is a substance user. The variable I) is perceived peer behaviors. The vector 2 contains a set of student. family, school. and HThe difference between the subjective measure and the objective measure of the peer behaviors implies that we cannot calculate the social multiplier effect from the size of endogenous effect (3). since we need to know how individuals formed their perceptions. Only with perceived peer behaviors. how- ever. can we confirm the existence of the endogenous effect. without assuming al.)sence of the contextual effect (i.e. '7 = (l). as discussed befm‘e. C}! to regional characteristics that may affect individual substance use. The variable .17 is an attribute of the group. which is ” schools" in this model. The determination of perceived behavior is specified as p = 91E(y|.r) + :92 + E(z|.r)93 + v. (3.17) I assume E (u|.r. z) = 0 and E (lll z) = 0. Inclusion of several measures of parental involvement in vector :3 such as participation in PTA meetings and many variables (85 total) that. may affect. a. youths substance use. makes the assumption E (u|.r..z) = 0 plausil.>le. There are still potential sources of omitted variable bias. For example. state anti- drug campaigns to reduce. teenager substance use is a. possible omitted variable that may affect both a respondents and his or her peers substance use (”correlated effect.” using Manski (1993)’s terminology). In this situation. the violation of E(u|.r. z) = 0 is likely to occur through E (ulr. z) = f (it). To reduce this possibility, I included many regional variables in 3 that characterize the county where the teenager lives (e.g., state cigarettes tax rate. beer tax rates. cormty-level poverty rate. county-level unemployment rate, the demographic composition of the county. and other characteristics). It is still fair to say. however. that the assiunption E ( ulr. :) = O can be violated. Thus, I will later relax this assumption and use fixed effect estimation. To ensure that the OLS is an unbiased and consistent estimator. I also assume E('u|.1‘. 2.1) = 0. (3.18) This means that the error term of the l)(;‘ll'd\'lOl"dl equation (a) is not correlated with the error term in the equation determining perceptions of peer behaviors (e). Under this assumption. applying OLS (3.16) renders an unbiased and consistent estimator of ,3. '7 " - ).'as-. and lecu‘e E(u|z.p..r) = E(ul.r. z.p(E(y|.r). z. e)) = E(ul.r. z. c) = 0. (3.19) 53 The cell average of 2 within a school is used instead of E(z|;r).15 Since 11 is heteroscedas- tic due to the binary dependent variable. the variance covariance matrix is calculated using the White formula. Since the respondents school i<.le.ntification number (ID) is not available in my data set. I create the (almost) school ID by matching the respondent’s county ID. school size and student / teacher ratio. 16 3.5.3 The results I report the results of the above model in Table 3—3. For cigarette smoking. the coefficient 0.221 means that. a 10 percentage point increase in the subjective perception of the peer smoking probability increases the probability of smoking by 2.2 percentage points. This estimate is statistically significant. For alcohol drinking. the estimated coefficient is 0.311. For marijuana usage. the OLS estimate is 0.229. The results clearly show the existence and statistical significance of peer effects. W hen a teenager’s perception of the 1')e.r('tentage of his/ her peers who use a substance increases by 10 percentage points. the prt.)bability that he/she will use the substance increases from 2.2 to 3.1 1.)er('rentage points. Although the difference in identification strategy prohibits me from serious comparison of the estimates. the estimated peer effect is comparable to the estimated effect of Gaviria. and Raphael (2000) for smoking and alcohol drinking (0.150 for smoking. 0.106 for alcohol drinking and 0.254 for drug 15Although this first step estimation may change the asymptotic distribution of the OLS estimat(_)rs. it is known that under the null of H0 : '7 = 0 the asymptotic distribution is not affected. lbSchool size is classified into five categories: 1-299. 300—499. 500-749. 750-999. and 1000+ students. Student/teacher ratio is classified into four categories: less than 14. 14-17, 18-21 and more than 22. The Bureau of Labor Statistics assigns these two variables to each respondent based on the confidential information held by BLS. Thus. when each school in a county is different in either school size or student / teacher ratio. I can identify all of the schools. If all of the schools in a county share both the same school size and student/teacher ratio. I just identify the county. In the worst-case scenario. it is to assume E (SlJf) is constant within a county. This mis-matching becomes serious if the variation of E (:31?) is huge within a county. However. the direction of bias in the estimator of 7' caused by this measurement error is not clear a. p‘r‘iori. since the measurement error is mean reverting (to the county mean) and not classical. The contextual effect that operates at the county level is, however. at least captured. use). The absence of the contextual effect (‘7 = 0) is not rejected through an F -test. In sun’nnary, the results of the estimation robustly show the existence of peer effects in the causal sense. .\Ioreover. due to the usage of perceived peer behavior as the key independent variable. the results show that peer effects work through the endogenous effect. This implies the existence of the ”social multiplier.” The causal interpretation critically depends on the assumption E(u|;1:, z. e) = 0. The multitude of variables in z (82 in total). l‘iowever, makes this assumption realistic.” 3.6 The (almost) school fixed effect estimation Although the previous section made the best effort to assure the assumption E (izlr. z, c) 0 is correct. the school attribute .r nonetheless may still contain some information which systematically predicts teenagers" substance use that the regional or school character- istics included in z fail to capture. In other words, there might be regional and school factors that encourage the teenagers’ substance use that are not observed. For example, suppose a cigarette shop is located just in front. of high school A. Moreover. suppose this ” unobservable” makes a student. in the high school 50 percentage points more likely to smoke. Then the assumption E(u|.r. :. if) = 0 is violated. because 0.5. if .r = A. E (11.3.14' = ~ - i l ) 0. ()therw1se. (3.20) This possibility is addressed through school fixed effect estimation, which can be repre- sented in the equation LU!) : jlhj + 3071+ E(:..‘r)j72 + (1(I)j +110. (3.21) Here 2'. is a subscript for an individual and j is a. subscript. for a school. The coefficient 72 is not identified, since E (2|;r) j is invariant within a school. The random variable {1(1)}- 17If the instrumental variables that are sufficiently correlated with p but not correlated with u exist. we can use Hausman (1978)‘s method to test endogeneity of p. In particular. specifying E(u.i.1:. :. I?) = pH, ”0 : p = 0 can be tested given E(ul;r. :) = 0. However. finding such IV is prohibitively difficult. captures the school specific unobservable which affects the teenagers substance usage, such as a cigarette shop at the school gate. Under the assumptions E(u|p. :.a..r) : 0 and E(a|p, 3.1?) = 0, both the random effects estimator and fixed effects estimator are consistent. However. only E (111 p. z. a) = 0 needs to be true for the consistency of the fixed effect estimator. This allows for a test of the assumption E ( a] 1). z. .r) = 0 using a Hausman test. Since the random effect estimator is not efficient due to heteroscedasticity, I use the robust form of the Hausman test introduced in \Vooldridge (2001). Although it is not rigorously justified, the single variable Hausman test roughly tests the null E(a|p, 2.51.“) = 0. Under this assumption, t.18 The results of the random, the estimator of the coefficient for p is roughly consisten fixed effect estimation and the Hausman tests appear in Table 3-4. All of the Hausman tests do not reject the null of E (0| p. z. .r) = 0. Thus I am in favor of using the random effect estimator because of its relatively efficient property. Peer effects for smoking, drinking and marijuana usage are 0.22. 0.31. and 0.23 respectively. These numbers are similar to the OLS results. This is probably due to the fact that the vector :5 already contains enough ii‘iformation to capture the school "‘ unobservable.” 3.7 The household fixed effect estimation using the sibling sample To reinforce the previous results. I estimate the household fixed effect model using sibling samples. One concern in the previous research (Evans, Oates, and Schwab (1992)) was the endogeneity of peer quality due to omitted household characteristics. Peer quality can be endogenous since parents who are willing to invest in their children send their children to a school with good peers. Parental care can also directly affect a child's behaviors. This problem can be avoided through controlling household unol.)served heterogeneity using the household fixed effect. estimation.19 1”This discussion is not exactly true as far as p is correlated with the other explanatory variables. 19The sibling method. however. does not necessarily solve the endogeneity. Since between household difference of outcomes and perceived peer behaviors are wiped away. the identification solely depends on I estimate the following model: 1],]- 23:31))j + $,‘j";1 ‘i' E(Zl;1'),‘j + (.‘j + Uij (3.22) where i is a subscript. for an individual. j is a subscript. for a household, and c is a household unobserved heterogeneity. The. same econometric discussion from the previous section applies. The results of random and fixed effects estin‘iation and the Hausman tests appear in Table 3-5. All of the Hausman tests reject the null of E(c|p. 2,1) = 0. Thus, the fixed effects estimator is preferred. These results are similar to the previous OLS results that appeared in Table 3—1. Peer effects for smoking, drinking and marijuana usage are 0.14, 0.27, and 0.21 respectively. The fixed effects coefficient estimates are smaller than those obtained from the random effects or OLS estimation due to the positive correlation of the household unobserved heterogeneity and the peer variable. Nevertheless, even after allowing for the correlation of the household heterogeneity and the peer variable through the fixed effect model, the estimated peer effects are practically and statistically significant. 3.8 Peer effects and the demographic groups Thus far, the existence of peer effects is a robust result. Next, it is interesting to inves- tigate the strength of peer effects within different demographic groups. With knowledge of the group in which the endogenous effects are strong, policy makers can effectively target the policy that is likely to (’liscourage youth substance use within that group, since he/ she can expect larger policy effects through the larger ”social multiplier” .20 To within household variation of outcomes and perceptions. If each sibling in a household has unol.)served characteristics that determine both substance use and perception of peer behaviors. the sibling estimator is still biased. If this within household unobservable plays a more important role than the between households unobservable. the sibling estimator can be more biased. However. this discussion is unlikely to apply in the context of substance use. See Bound and Solon (1999) and Neumark (1999) for possible biases in the sibling estimator of the return to education. 20This discussion assumes. though. that sensitivity to the policy (a part of 1}) is the same across groups. 57 estimate the difference in the endogeinms effects across groups, I assume the following model in which the strength of peer effects may depend on the different demographic groups: y = (i + .311) + p * 21.32 + :7 + E(:l.1‘)'l] + u. (3.23) ‘) . . “I Since the exogeneity where 21 is the part. of 3 that defines demographic groups. of p was not rejected in the school and household fixed effects estimation, I assume E(ulp, 3,1“) = 0. Under this assumption, OLS is an unbiased estimator. The results of the regressions appear in Table 3-6. Some of the estimated coefficients on the interaction terms are statistically significant and the effects are not. negligible. As for the marijuana usage, females are less likely than males to be affected by their peers. For the other two substances. a gender difference was not found. Fewer peer effects were found among black teenagers. For smoking behavior, peer effects among black teenagers are one-third of that. found anmng white teenagers. The smaller peer effects among black teenagers are statistically different from zero for all substances. Hispanics are also less vulnerable to peer pressure. An expectation is that. minority teenagers might not obtain as much utility as non-minority teenagers from imitating each other. Teenagers with both biological parents are less likely to be affected by their peers in their smoking and marijuana uses. As for drinking, the coefficient on the interaction term of the peer variable and ”both biological parents" is not statistically significant. The first two results may imply that. the teenager who does not have both biological parents present is more likely to depend on his peers to form his behavior. This result is consistent with Steinberg (1987):}. The result was not. obtained in Caviria and Raphael (2000), probably because the peer variable only interacts with the single parent dummy, not with the race dummies. 2‘1 also estimated the endogemms effect using a sub—sample of demographic groups. The results obtained were qualitatively the same as the results obtained in this section. 3.9 Conclusion The estimation of peer effects on substance usage through perceived peer behaviors shows the existence of significant peer effects. “hen the perceived peer substance use is doubled, the probability that. a teenager will use substances increases by forty to sixty percentage points. .\I<:)reover, the endogenous effect is found to be more important than the contextual effect when explaining the peer effects on youth substance use. This finding implies that. current peer behaviors, rather than peer backgrounds, determine individual behaviors. Thus, if some exogmious shock reduces a groups substance use, this reduction propagates to other groups of youths through the endogenous effect. Hence, policy makers can expect. a "social multiplier” effect in policies that discourage youth substance use. In my model, the endogenous effect is identified when perceived peer behaviors are exogenous. To assure this exogeneity assumption, I used a rich set of controls consisting of parent, neighborlmod and school cltaracteristics. ;\I(;)reover, the robustness of the results is confirmed through the school and household fixed effect estimations. we also find that. the strength of peer effect. depends on the demographic group to which a teenager belongs. Peer effect is found to be large among white teenagers and teenagers without both biological parents. Although this paper finds a robust peer effect. this study does not shed enough light on the mechanism of peer effect itself. Thus this study is still a ” reduced form” study of peer effects. More direct research on the mechanism of peer effect is left for future research. Also. peer effects on the several other socioeccmomic behaviors such as criminal or sexual activities are the topics for the future research. I hope the robust findings in this study stimulate further investigation in this field. 59 Bibliography AARONSON. D. (1998): "Using sibling data to estimate the impact of neighborhoods on children’s educational (:)utco111cs." Journal of Human Resources. 33(1). 915—946. AKERLOF. G. A. (1997): "Social distance and social decisions," Econometrica, 65(5), 1005—1027. ARABSHEII'MNI. G.. D. l)ICl\1EZA. J. I\IALO.\IEY. AND B. PEARSON (2000): “And a Vision Appeared unto them of a Great Profit: Evidence of Self-Deception among the Self-Employed,” Economics Lcllcrs. 67. 3Al~4l1. BECKER, G. S. (1996): Accounting for tastcschap. 11. Harvard University Press. BENDER, K. A., AND P. J. SLOANE (1998): "Job satisfaction. trade unions, and exit-voice revisited." Imlastrlal and Labor Relations Review. 51(2). 222—240. BIDDLE. J. (1991): "A bandwagon effect in personalized license plates‘?,” Economic Inquiry, 29(2). 375 388. BLANC‘HFLOWER. D. 0.. AND A. J. OswALI) (1998): "What makes an entreprenem‘.’,"‘ Journal of Labor Economics. 16(1). '26 (if). BQRJAS. G. .l. (1979): ".lol)satisfaction. wages and unions,“ Journal of Human Re— sources, 11(1). 2140. BORJAS, G. J.. AND S. G. BRONARS (1989): "Consumer (liscrimiiration and self- emj')loyment." Journal of Political Economy. 97(3). 581 605. 60 BOUND, J., AND G. SOLON (1999): "Double trouble: on the value of twin-based estimation of the return to schooling," Economics of Education Review, 18, 169—182. BROCKHAUS. R. H.. AND P. S. HORWITZ (1986): "The Psychology of Entreprcnem',” in The Art of Science of Entrcprcnc-urshlp, ed. by D. L. Sexton, and R. W. Smilor. Ballinger, .\lassacluisetts. CARI). D. (2000): "The Causal Effect of Education on Earnings," in Handbook of Labor Economics, ed. by O. C. Ashenfclter, and D. Card, vol. 3A, chap. 30. Elsevier, New York. CARRlNGTON. W. J.. K. MCCUE. AND B. PIERCE (1996): “The role of em- ployer/emj)loyee interactions in labor market cycles: Evidence from the self- employed.” Journal of Labor Economics. 11(1). 5717602. CASE. A. C.. AND L. F. KATZ (1991): "The company you keep: The effects of fam- ily and neighborhood on Disadvantaged youths.” Discussion paper, NBER “forking Paper Series May. CLARK, A., Y. Cl£()R(;l£Ll.IS. AND P. SANFEY (1998): “.lol) satisfaction, wage changes, and quits: Evidence from Germany." Rcsearclz. in Labor Economics, 17, 95-121. CLARK. A. E. (2001): "What really matters in a job‘.’ Hedonic measurement using quit data,” Labour Economics. CLARK, A. E.. AND A. J. OSWALD (1991): "linhappiness and L'nen1}.)loyment,"’ Eco- nomic Journal. 101(121), 618‘ 59. (1996): "Satisfaction and cmnparison income,” Journal of Public Economics. 61, 359 381. Cooeua. A. C.. C. Y. Woo. AND W. C. DUNKELBICRC (1988): "Entrepreneurs’ Percieved Chances for Success." Journal of Business Venturing. 3. 97—108. (51 CRANE. J. (1991): "The epidemic theory of ghettos and neighborhood effects on drop- ping out. and teenage childbearing.” American. Journal of Sociology, 96(5), 1226-1259. DUNN. T., AND D. HOLTz-EAKIN (2000): “Financial capital, human capital, and the transition to self—employment: Evidence from intergeinérational links,” Journal of Labor Economics, 18(2). 282 305. EVANS, D. 8., AND B. JOV’ANOVIC (1989): "An Estimated Model of Entrepreneurial Choice under Liquidity Constraints," Journal of Political Economy. 97(1), 808—27. EVANS, D. 8.. AND L. S. LEIGHTON (1989): "Some empirical aspects of entrepreneur- ship,” Aimrrican Economic Rccicw, 79(3), 519 535. EVANS. W. N., W. E. OATES. AND R. M. SCHWAB (1992): "Measuring peer group effects: A study of teenager behavior." Journal of Political Economy, 100(5), 966~991. FAIRIJE, R. \V., AND B. D. MEYER (1996): "Ethnic and racial self-eniployment (lif- ferences and possible exj’)la.nations." Journal of Human. Rmources, 31(4), 757—793. FREEMAN, R. B. (1978): “Job satisfaction as an economic variable," American Eco- nomic. Rcuicun 68. 135—111. GAVIRIA. A.. AND S. RAPHAEL (20(10): "School based peer effects and juvenile behav- ior,” Forthcoming in Review of Economics and Statistics. CRUISER. J. (2000): “Youth smoking in the 1.7.8.: Prices and policies,” Discussion paper, NBER \Vorking Paper Series 7506. HAMERMESII, D. S. (1977): "Economic aspects of job satisfation,"7 in Essays in Labor Market and Population Analysis, ed. by Ashenfelter, and Oates. \Viley. (2001): "The Changing Distribution of .101) Satisfaction,” Journal of Human Resources, 36(1). 1‘30. HAMILTON, B. H. (2000): "Does Entrepreneurship Pay? An Empirical Analysis of the Return to Self—E1nployment," Journal of Political Economy, 108(3), 601 631. HAUSMAN. J. A. (1978): “Specification tests in econometrics," E(0110nltl7u(l 46(6), 1251—1271. HEC‘KMAN J. J. (1976): ".A life—cycle model of earnings. learning, and consumption,” Journal of Political Economy, 811(1). 811- 11. HOUT, l\'I., AND H. S. ROSEN (1999): "Self-employment, family background, and race,” Discussion Paper 7-133, N BER. JOULFAIAN, D., AND M. RIDER (1998): "Differential taxation and tax evasion by small business Vational T111 Jouinal. 51(1), "75-687. JOVANOVIC, B. (1979): "Job matching and the theon of tulnm e1 Journal of Political EconomyS 7(5), 972 990. KANBUR, S. M. R. (1982): "Entrepreneurial risk taking. inequality. and public policy: An application of inequality 1lecomposition analysis to the general equilibrium effects of progressive taxation." Journal of Political Economy, 90(1), 1*21. KlHLSTROM. R. E.. AND J.-J. LAI‘EONT (1979): A general equilibrium en- trepreneurial theory of firm formation based on risk aversion." Journal of Political Economy, 81(1). 719718. KR \SHINSKI H. (2000):) Sele111plm111eIIt and “(we Inequalm P1111(etonLnI\e1s1t\ Mimeo. LAZEAR. E. P. (1979): "\Vhy is there 111a.ndatory i'eti1'(:‘111elrt‘?.” Journal of Political Economy. 87(6), 1261- 128 1. LAZEAR. E. R, AND R. L. MOORE (1981): "Incentives. P1'1_)ductivity. and Labor Contracts," Quarterly Journal of Economics. 99. 275—295. 63 LEIGH. D. E. (1979): "Do Union Members Receive Compensating \V’age Differential - Note.” American. Economic Review, 11(1), 21—410. LINDII, T.. AND H. ORLSSON (1996): "Self-employment and windfall gains: Evidence from the Swedish lottery.” Economic Journal. 106(139), 1515-1526. LOS-ANGELES-TIMES (1999): "Smoking among youths continue to rise," January 8. LUCAS, R. E. (1988): "On the .\11_‘cln~1.11ics of Economic Development," Journal of Mon- etary Economics. 22(1). 3 12. MANSER. M. E., AND G. PICOT (1999): "The Role of Self-E1111:)loy1nent in US. and Canadian .1111) Growth." Monthly Labor Rcci1—‘ttc. 122(1), 107725. MANSKI. C. (2000): “Econtnnic Analysis of Social Interactions." Journal of Economic Pcrspcctim-f. 68(3), 115 36. MANSKI. C. F. (1993): "Identification of endogenous social effects: The reflection problem," Roiiicin of Economic Studics, 60(3), 531 512. MCCORMICK. E. J.. AND D. ILCEN (1985): Industrial and Organizational Psychology. Prentice Hall. New Jersey. eighth edn. MOSKOWITZ. T. J.. AND A. VISSING-JORGENSEN (2001): "The Private Equity Pre- mium Puzzle.” Graduate School of Business. University of Chicago. MUNDLAK, Y. (1978): "On the Pooling of Time Series and Cross Section Data.” Econo- metrica. 416(1). 691—85. MUNSIII, K. (2000): "Social leaning in a. heterogeiIeous 1')opulation: Technology diffu- sion in the Indian Green Revolution." .\Ia.nuscript. NEUMARK. D. (1999): "Biases in twin estimates of the return to schooling.” Economics of Education. Recicut 18. 113-118. 64 NEUMARK. D.. AND A. POSTLEWAITE (1998): "Relative income concerns and the rise in 111arried women’s employment." Journal of Public Economics, 70. 157 183. NORTON. E. C.. R. C. LINDROOTII. AND S. T. ENNETT (19.98): “Controlling for the endogeneity of peer substance use on adolescent alcohol and tobacco use," Hcaltli Economics, 7. 139—7153. SACERDOTE, B. (2000): "Peer effects with random assignment: Results for Dartmouth roommates." Discussion paper, NBER \V'orking Paper Series 7-169. SALOP, J .. AND S. SALOP (1976): "Self-selection and turnover in the labor market," Quatcrly Journal of Economics, 90(1). 619627. SCHILLER, B. R.. AND P. E. CRIM'SON (1997): "Entrepreneurial origins: A longitu- dinal inquiry." Economic Inquiry. 35(3), 52377531. SIECEL, I... AND I. M. LANE (1987): Pcrsonncl and Organizational Psychology. Irwin, Illinois. STEINBERG. L. (1987): "Single parents. step parents, and the susceptibility of adoles- cents to antisocial peer pressure," (71 1111 Dccclopmcnt, 58, 269275. \A'fOOLDRIImE, J. M. (2001): Economctric Analysis of Cross Section and Panel Data. MIT press. Cambridge. MA. Table 1-1: The replication of the Lazear and Moore (l984)’s finding and the risk of self-employed workers. (1 ) (2) (3) (4) (5) (6) Model OLS F.E. OLS F.E. OLS F.E. Dependent variable In wage In wage Residual2 Residualz Residual2 Residual2 0H2) 0f(2) of(2) of(2) education 0.070 — — — — — (0.003) experience 0.063 0.1 12 -— — — — (0.005) (0.008) experience-7100 -0120 -0.187 .. - _ _ (0.026) (0.019) tenure 0.052 0.032 — — — — (0.004) (0.003) tenure-7100 -0250 0198 _ _ _ _ (0.029) (0.020) self- employed 0.032 0.086 0.237 0.189 0.201 0.172 (0.033) (0.018) (0.041) (0.038) (0.056) (0.046) self X (educ- educ ) 70007 7 7 7 7 7 (0.01 1) self x (8» ex) -0.043 -0.043 — — — — (0.020) (0.012) sele(ex2- ex2)/100 0.193 0.146 — — — — (0.088) (0.049) self x (1611- ten) -0.054 -0.03l — — — — (0.016) (0.009) selfx (tenJ- ten2 )/ I ()0 0243 0'228 7 7 7 7 (0.100) (0.057) (self- W): _ — — — 0.133 0.059 (0.092) (0.090) Constant 5.566 — 0.147 — 0.145 — (0.058) (0.009) (0.088) Observations 23887 23887 23887 23887 23887 23887 Number of individuals — 2715 — 2715 — 2715 R2 0.31 _ _ _ _ _ Wage of SE — Wage of SW 0 year experience and 0.343 0.357 0 year tenure (0.107) (0.061) 5 years experience and -0.034 0.080 5 year tenure (0.059) (0.032) 10 years experience and -0.l92 -0.01 l 10 year tenure (0.062) (0.035) Notes: , l. Standard errors are in parenthesis for coefficient estimates. For OLS estimates, standard errors are corrected for the panel clustering. 2. Dependent variable in the each regression (3), (4), (5) and (6) is the squared idiosyncratic residual of regression (2 ). 3. Year dummies are included but coefficients are not reported. 4. Wage of SE - Wage of SW is evaluated at the sample mean of education that is 13.107. 66 Table l-2: School Enrollment by Sectors (1988-1998) Salary and Self Total Wage Employed Enrollment Rate 9.01 5.60 8.67 N 16021 1804 17825 Notes: 1. The workers who were enrolled in school at all between last interview date and cunent interview date are classified as “enrolled.” This does not necessarily imply ' cunent enrollment. 2. Only observations afier year 1988 are included in the sample. Youngest respondents are 24 in 1988. Table 1-3: Linear Probability Model of School Enrollment Dependent Variable: Enrolled in School (Yes=l, No=0) (1) (2) Mean of Dependent Variable 0.087 0.087 Method of Estimation OLS F.E. Self-Employed -0.024 -0.019 (0.008) (0.009) Education 0.026 - (0.002) Experience -0.015 -0.040 (0.004) (0.006) Experience”) 0.045 0.081 (0.015) (0.013) Tenure -0.012 -0.007 (0.002) (0.002) remit-e3 0.055 0.039 (0.011) (0.012) Constant -0. 138 0.508 (0.030) (0.080) Observations 17825 1 7825 R-squared 0.07 0.02 Number of 1D — 2634 Note: The same note applies as Table 2. Standard errors are in parenthesis for coefficient estimates. For OLS estimates, standard errors are corrected for the panel clustering. 67 Table 1-4: Training Participation by Type of Training and Cost Bearing (1988-1998) Panel A: Training Participation by Type of Trainings Salary Self— Total and Employed Wage Not Participate in Training 81.09 90.74 82.06 On—the-job Training 1 1.67 3.61 10.85 Apprenticeship Program 0.75 0.28 0.70 Formal Company Training Run by Employer or Military Training 7.35 1.83 6.79 Seminars or Training Programs at Work Not Run by Employer 3.57 1.50 3.36 Off-the -job Training 6.20 5.00 6.07 Business School 0.29 0.33 0.29 Vocational or Technical Institute 1.45 0.67 1.38 Correspondence Course 0.52 0.28 0.49 Seminars or Training Programs outside of Work 3.71 3.55 3.69 Vocational Rehabilitation Center 0.14 0.1 l 0.13 Government Training Program 0.09 0.06 0.09 Other 1.05 0.67 1.01 N 16014 1804 17818 Panel B: Training Participation by Type of Cost Bearing Salary Self- Total and Employed Wage Not Participate in Training 81.09 90.74 82.06 Training Cost Paid by Self or Family 1.57 3.88 1.81 Employer 15.87 3.60 14.63 Job Training Program Act 0.14 0.1 l 0.14 Trade Adjustment Act 0.02 0.06 0.02 Job Corps Program 0.01 0.00 0.01 Work Incentive Program 0.01 0.00 0.01 Veteran’s Administration 0.02 0.06 0.03 Vocational Rehabilitation 0.09 0.00 0.08 Other 1.17 1.55 1.21 N 16014 1804 17818 Notes: 1. The results are based on most recent participating training program that starts between the last interview date and the current interview date. 2. Among types of training, “Business School,” “Vocational or Technical Institute,” “Correspondence Course,” “Seminars or Training Programs outside of Work” and “Vocational Rehabilitation Center” are classified as off-the- job training. 3. Trainings paid by “Self or Family” and govemment programs are classified as “own cost training.” 68 Table 1.5; Linear Probability Models of Training Participation (l) (2) (3) (4) (5) (6) (7) Sample SE and SW SE and SW SW Only Dep. Var. Training Ofi‘J’I‘ Training Ofi-JT Own/Gov. Paid Mean 0.179 0.061 0.179 0.061 0.033 Method of OLS F.E. OLS r15. OLS OLS OLS Estimation SE -0097 -0050 -0012 -0010 - - — (0.009) (0.013) (0.006) (0.009) Future Self — _ _ - -0057 -0.001 0.014 (0.012) (0.008) (0.006) Education 0.024 — 0.008 — 0.024 0.006 0.000 (0.002) (0.001) (0.002) (0.001) (0.001) Experience -0.003 0.000 -0.003 -0.005 -0.003 -0.005 -0.006 (0.004) (0.009) (0.003) (0.006) (0.005) (0.003) (0.003 ) Experience) 0.023 . 0.010 0.020 0.016 0.020 0.025 0.019 (0.018) (0.019) (0.011) (0.013) (0.025) (0.016) (0.012) Tenure 0.006 -0006 -0.001 -0002 0.008 -0.001 -0005 (0.003) (0.003) (0.002) (0.002) (0.003) (0.002) (0.001) Tenure’ -0.038 0.025 0.009 0.022 -0049 0.014 0.022 (0.018) (0.017) (0.012) (0.012) (0.025) (0.016) (0.008) Constant -0.148 0.195 -0035 0.117 -0103 0.000 0.095 (0.032) (0.120) (0.018) (0.079) (0.037) (0.021) (0.017) Observations 17818 17818 17818 17818 13671 13671 13671 R-squared 0.04 0.01 0.01 0.00 0.03 0.01 0.01 Number of [D - 2634 - 2634 - - - Note: The same note applies as Table 2. Standard errors are in parenthesis for coefficient estimates. For OLS estimates, standard errors are corrected for the panel clustering 69 Figure 1-1: Life Time Utility of Each Career Path and Initial Human Capital Life Time Utility (1‘) A 315— SEW”) SW - SE (VH) SW 7 SW (12,.) / > Initial Human Capital (h,,) The parameter of absolute risk aversion (y) is given Note: SE: Self-Employed workers; SW: Salary and Wage workers. Figure 1-2: Distribution of test scores among SE and SW workers —a— AFQT Score for SW Workers ——9— AFQT Score for SE Workers .015 A .01 " l T 100 0—4 . 50 Percentile of AFQT Score in 1989 Note: Bandwidth = 6 and Epanechnikov kernel was used to estimate the kernel density. The distribution of percentile ranges from -5 to 105 because of the bandwidth = 6, actual distribution of the percentile of AFQT89 ranges from 1 to 99. Workers who were both SE and SW over their careers contribute to both populations on a career-weighted average. 70 Table 2-1: Sample Construction Total Salary Self- Wage employed workers workers Original NLSY79 1985-1998 152232 White 90120 Male 45480 Employed + out of school 24756 Work in private. govemment and self-employed 24580 Valid answer for job satisfaction: Sample ( 1) 24533 22095 2438 Employed + out of school for two consecutive interviews 19893 Valid class + tenure variables 19402 Work in private. govemmcnt and self-employed for two 19298 consecutive interviews Valid answer for job satisfaction: Sample (2) 19222 17265 1957 Sample (2) + lagged demographic variables are available: 13889 12512 1377 mple (3) Sample (1) + valid covariate + more than 2 years of 20454 18585 1869 observation: mple (4) Hourly wage/eamings are between 1 percentile and 99 20065 18328 1737 percentile: amateur Note: Tenure variable is used to identify job change. Sample ( 1) is used in the analysis of Table 2-3. Sample (2) is used in the analysis of Table 2-4. Sample (3) is used in regression analysis. 71 m. Table 2-2: OLS Regression Coefficients Dependent variable: log hourly wage Sample: White Male (Sample (4)) (1) (2) OLS Fixed Effects Self-employment -0.01 1 0.022 (0.030) (0.016) Married-Spouse present 0.093 0.056 (0.014) (0.010) Education 0.078 - (0.003) Experience 0.047 0. 1 09 (0.006) (0.010) Experience: / 100 -0.068 0.115 (0.028) (0.021 ) Tenure 0.046 0.029 (0.004) (0.003) Tenure2 / 100 .0223 -0.186 (0.028) (0.021 ) Self-employment x Married - Spouse Present -0.059 -0.000 (0.061 ) ( 0.028) Self-employment x Education -0.0()8 - (0.012) Self-employment x Experience -0.020 -().024 (0.023) (0.013) Self-employment x Experience2 / 100 0.1 15 0.080 (0.095) (0.053) Self-employment x Tenure -0.050 -0.027 (0.016) (0.009) Self—employment x Tenure2 / 100 0.221 0.194 (0.102) (0.058) Constant 5.162 5.923 (0.044) (0.044) Observations 20454 20454 R—squared 0.30 0.26 Number of individuals 2661 2661 Wage of Self-Employed — Wage of Wage and Salaried Workers No Experience and No Tenure 0.184 0.209 (0.128) (0.071) 5 Year Experience and 5 Year Tenure -0.084 0.027 (0.064) (0.036) 10 Year Experience and 5 Year Tenure -0.184 -0.018 (0.064) (0.037) Standard errors robust against panel clustering are in parentheses for OLS estimates. Standard errors are in parenthesis for F .E. estimates. Table 2-3: Job Satisfaction among Workers Panel A Job satisfaction among salary/wage workers Sample: White Male (Sample (1)) year Like very Like fairly Dislike Dislike very Total much well somewhat much observation Row % Row % Row °/o Row “/6 85 45.97 43.27 8.65 2.1 1 1849 86 43.25 48.13 6.13 2.49 1926 87 41.95 50.56 5.97 1.52 2043 88 41.28 48.82 8.20 1.70 2122 89 43.59 46.89 7.43 2.09 2154 90 41.58 49.84 6.23 2.35 2167 91 43.17 47.58 7.41 1.84 1633 92 43.00 48.49 6.51 1.99 1658 93 42.19 48.85 7.69 1.27 1652 94 41.51 50.60 6.39 1.51 1660 96 43.1 1 48.93 6.55 1.41 1633 98 46.12 46.62 5.32 1.94 1598 Total 43.00 48.24 6.89 1.87 22095 Panel B: Job Satisfaction among Self-Employed Workers year Like very Like fairly Dislike Dislike very Total much well somewhat much observation Row “/6 Row "/6 Row % Row % 85 72.61 22.29 3.18 1.91 157 86 63.07 31.25 1.70 3.98 176 87 59.90 36.04 2.54 1.52 197 88 64.71 31.22 3.17 0.90 221 89 64.38 33.48 1.72 0.43 233 90 65.07 31.00 3.06 0.87 229 91 70.05 26.40 2.54 1.02 197 92 63.26 33.95 2.33 0.47 215 93 63.59 33.98 1.94 0.49 206 94 62.56 33.33 3.08 1.03 195 96 74.52 22.60 2.40 0.48 208 98 65.69 29.41 3.92 0.98 204 Total 65.67 30.60 2.63 1.1 1 2438 73 Table 2-4: Probability to Change Job by the next Interview Sample: Sample (3) Job satisfaction in the previous interview Salaried and Self-Employed Mean Wage Worker Workers Like very much 0.148 0.108 0.143 Like fairly well 0.177 0.207 0.179 Dislike somewhat 0.256 0.383 0.263 Dislike very much 0.388 0.266 0.379 Total 0.173 0.153 0.171 N 12512 1377 13889 Note: Sample is constructed to be consistent with the sample used for the job change regression. Table 2-5: Job Change Probit Model Dependent variable: Job Change between t and t—l(’Yes= l; No=0) Sample: Sample (3) ( 1) (2) Lagged job satisfaction: Dislike very much 0.764 0.734 (0.098) (0.099) Laged job satisfaction: Dislike somewhat 0.434 0.432 (0.051) (0.052) Lagged job satisfaction: Like somewhat 0.149 0.145 (0.029) (0.030) Self Employment -0.013 0.006 (0.062) (0.064) Self Employment x Dislike very much , —0.627 -0.567 (0.308) (0.291) Self Employment x Dislike somewhat -0.123 -0.103 (0.284) (0.285) Self Employment x Like somewhat 0.135 0.130 (0.099) (0.102) Laggcd education -().017 -0.005 (0.005) (0.006) Lagged experience ' -0.012 -0.005 (0.004) (0.004) lagged tenure 41.046 -0.047 (0.004) (0.004) Laggcd log wage -0. 135 (0.026) Constant -0.545 0.164 (0.080) (0.166) Log Likelihood .6163 ~5979 Pseudo R2 0.029 0.033 °/o Conectly Predicted 82.92 81.78 Observations 13889 13592 Marginal effects of Probit model is presented. Robust standard errors against panel clustering are in parentheses. Lagged job satisfaction: like very much is the omitted category. 74 Table 2-6: Transition of Job Satisfaction Associated with Job Change (Percentage) (SE: Self-Employment Job. SW: Salary and Wage Job) Sample: White Male (Sample (2)) Like Like Dislike Dislike Very Much Somewhat Fairly Well Very Much W W Egel A: Job Stayers 1 /war - l /w N 13910 Like Very Much 30.17 10.34 0.96 0.19 Like Somewhat 13.11 33.26 2.95 0.52 Dislike Fairly Well 1.03 3.79 1.62 0.29 Dislike Very Much 0.31 0.74 0.45 0.27 Egel B; Job Stags (SE-SE) N=1327 Like Very Much 54.56 9.87 0.38 0.15 Like Somewhat 13.49 17.71 0.75 0.30 Dislike Fairly Well 0.53 0.75 0.60 0.15 Dislike Very Much 0.08 0.30 0.23 0.15 W /w - l /w N=2838 Like Very Much 21.71 21.04 3.84 1.41 Like Somewhat 12.37 25.55 5.25 1.27 Dislike Fairly Well 1.30 2.85 1.30 0.42 Dislike Very Much 0.42 0.81 0.25 0.21 ' ' vi us (ear Job gjhanger ‘- =4 Like Very Much 51.22 12.20 2.44 0.00 Like Somewhat 4.88 24.39 2.44 0.00 Dislike Fairly Well 0.00 0.00 0.00 2.44 Dislike Very Much - - - - wa / ala - N= 89 Like Very Much 40.41 23.09 3.23 0.85 Like Somewhat 7.13 17.15 2.89 0.68 Dislike Fairly Well 1.19 1.87 0.51 0.17 Dislike Very Much 0.17 0.51 0.17 0.00 Eanel F: 19b g; hanger (SE-glm/wage) [11:51 7 Like Very Much 39.26 12.19 1.74 0.39 Like Somewhat 18.76 20.50 0.58 0.58 Dislike Fairly Well 2.32 1.35 0.39 0.19 Dislike Very Much 0.97 0.77 0.00 0.00 75 Table 2-7: The Results of Ordered Probit Estimation Dependent variable: “Like very much”=4. “like somewhat”=3, “dislike sorrewhat“=2, “dislike very much =1” Sample: Sample (5) (l) (2) (3) (4) (5) (6) Pooled Pooled ”Fixed "Fixed "Fixed "Fixed Ordered Ordered Effect" Effect" Effect" Effect" Probit Probit Ordered Ordered Ordered Ordered Probit Probit Probit Probit with with with Feedback Feedback Feedback Self Employed 0.486 0.489 0.4 1 6 0.43 l , 0.43 1 0.428 (0.051) (0.051) (0.047) (0.054) (0.054) (0.054) Log wage 0.248 0.232 0.341 0.315 0.315 0.315 (0.024) (0.031) (0.033) (0.036) (0.036) (0.036) Married 0.044 -0.067 -0.071 -0.071 -0.071 (0.026) (0.026) (0.029) (0.029) (0.029) Education 0.023 0.042 0.039 0.039 0.039 (0.006) (0.030) (0.038) (0.038) (0.038) Experience 0.014 0.001 0.004 0.004 0.004 (0.005) (0.009) (0.01 1) (0.01 1) (0.01 1) Mean(self employed) 0.129 0.143 0.156 0.157 (0.086) (0.096) (0.099) (0.102) Mean(log wage) -0.171 -0.127 -0.127 -0. 127 (0.049) (0.052) (0.052) (0.052) Mean (married) 0.184 0.185 0.185 0.185 (0.047) (0.050) (0.050) (0.050) Mean( education) -0.0 16 -0.015 -0.014 -0.015 (0.031 ) (0.039) (0.039) (0.039) Mean(experience ) 0.014 0.010 0.010 0.010 (0.009) (0.011) (0.011) (0.011) Fe ck ' s lace Within 1 year 0033 (0.046) Within 2 years 0042 (0.046) Within 3 years 0032 (0.047) Year dummies? Yes Yes Yes Yes Yes Yes 3rd Cut Point (113) -().377 -0.238 -0.533 -0.478 -0.481 -0.480 (0.165) (0.188) (0.241) (0.258) (0.258) (0.258) 2nd Cut Point (112) 0.360 0.501 0.208 0.282 0.279 0.280 (0.162) (0.185) (0.238) (0.256) (0.256) (0.256) lst Cut Point (111) 1.891 2.038 1.747 1.823 1.820 1.821 (0.164) (0.188) (0.240) (0.258) (0.258) (0.258) Mean of predicted latent 1.757 1.903 1.613 1.688 1.685 1.686 variable (u) Monetary Value of 7.104 8.182 3.387 3.930 3.928 3.893 Self-Employment status (2.596) (4.277) (0.697) (1.045) (1.125) (1.095) Observations 20065 20065 20065 17344 1 7344 17344 _L-ogLikelihood -19159 -19094 -19066 -16460 -16460 -16460 Note: Panel clustering robust standard errors are in parenthesis for pooled ordered probit estimates. Standard errors are in parenthesis for "Fixed Effects" effect ordered probit estimates. Married dummy is one if married and spouse present. zero otherwise. Monetary value of self-employment status is calculated by Exp[ (coefficient for self-employment)/(coefflcient for log wage)]. Standard error for this value is calculated through bootstrapping of500 repetitions. 76 Table 3-1: Sample Construction Average incidence (Non weighted) N Smoking Drinking Marijuana use (1) “711019 sample 8984 0.162 0.185 0.086 (2) All outcomes are ax ailable 8940 0.161 0.185 0.085 (3) Demographic variables are available 8851 0.161 0.185 0.086 (4) Relationships with parent are available 8833 0.161 0.185 0.086 (5) All peer variables are available 8518 0.165 0.190 0.088 (6) School characteristics are available 7498 0.166 0.192 0091 (7) Grade in school is available 7495 0.166 0.192 0.091 (8) Parent‘s 11(1C available 7491 0.166 0.192 0.091 (9) Variables from parent questionnaire are available 6615 0.168 (“93 0.092 (10) Proxy variables are available‘ 6356 0.168 0.195 0.092 (Basic analysis sample) (1 1) Within school duplicationoccurs’” 6312 0.167 0.196 0092 (School fixed effect analysis sample) (12) Siblings data are available'“ 2458 0.170 0.192 0.093 (Siblingirxed effect analysis sample) Note: *Proxy variables for school quality (If experience threat. if something stolen in school, Feel safe in school.) and neighborhood quality (if any gang in neighborhood). "Since school identification number is not available. quasi-school id, which is made out of county dummy. school size. and student -teacher ratio are used as quasi-school id. Bureau oflabor statistics assigns last two variables based on the school id number. "*Siblings are detennined by the identical household id. 77 Table 3-2: Descriptive Statistics of Substance Use and Srbjective Measure of Peer's Behavior by Grade Grade Smoke last Peer who Drunk last Peer who Use Peer who Number of 30 days smoke 30 days get drunk marijuana use illegal Observation (Subjective) (Subjective) last 30 days drug (Subjective) 4 0 0 0 0 0 0 2 5 0.044 0.1 18 0.041 0.077 0 0.068 37 (0.043) (0.055) (0.033) (0.039) (0.038) 6 0.058 0.147 0.039 0.037 0.018 0.084 426 (0.013) (0.013) (0.01 1) (0.007) (0.007) (0.010) 7 0.097 0.273 0.084 0.1 1 1 0.036 0.174 1294 (0.009) (0.009) (0.009) (0.007) (0.006) (0.008) 8 0.153 0.375 0.146 0.202 0.067 0.257 1319 (0.01 1) (0.009) (0.01 1) (0.008) (0.008) (0.009) 9 0.238 0.529 0.269 0.389 0.145 0.426 1416 (0.013) (0.008) (0.013) (0.009) (0.011) (0.010) 10 0.260 0.553 0.329 0.481 0.142 0.476 1204 (0.014) (0.008) (0.015) (0.009) (0.01 1) (0.010) 1 1 0.260 0.553 0.329 0.481 0.154 0.474 624 (0.014) (0.008) (0.015) (0.009) (0.016) (0.010) 12 0.291 0.555 0.371 0.513 0.154 0.487 34 (0.021 ) (0.01 l ) (0.022) (0.012) (0.016) (0.060) Total 0.190 0.427 0.214 0.303 0.099 0.333 6356 (0.006) (0.004) (0.006) (0.004) (0.004) (0.004) Note. 1. All statistics are calculated using sampling weight. 2. Standard errors of mean are in parenthesis. 78 Table 3-3: OLS Estimates of Incidence of Substance Use ( 1) (2) (3) Dependent variable " ‘ "' ‘ ‘ 111W mcidsnmtmamuana smgking in last 30 @ys rinkin in l 5 use in last 30 days Method of estimation OLS OLS OLS Pgr ( Fractign) Peer smoke 0.218 (0.017) Peer drunk 0.309 (0.020) Peer illegal drug 0.228 (0.015) 2 (Other control Yes Yes Yes variables) E[z 1x] Yes Yes Yes (Contextual Effect) F-statistics for 1.04 1.01 1.16 Contextual effect (0.379) (0.447) (0.163) R2 0.145 . 0.158 0.128 Sample size 6356 6356 6356 Note. 1. The vector zcontains following variables: 2 (Independent Variables): Dummies if respondent lives with mother, father, biological mother. biological father, foster mother or foster father. Female dummy, age, black dummy, other minority dummy, Hispanic dummy, school grade dummies (grade 5 — grade 12), [Middle/junior high] school and high school dummies, catholic school dummy, private school dummy, student/teacher ratio category dummies (3 categories), School size category dummies (3 categories), Census regional dummies, urban dummy and proxy variables for unobserved school characteristics ( lf experience thread, if something stolen in school, Feel safe in school), parents back ground variables (parent born in US, parent speak a language other than English in home. parent was with both biological parents at age of 14), last year‘s household income category dummies (less than $20,000, 320,001-340,000. $40.001-S60,000. 560,001-880,000, more than $80,001, and household income not available). household size. number of household member less than age 18 and less than 6, proxy variables for parent’s involvement in education (Often or sometimes participate in PTA activity, oflen or sometimes volunteer in school education) dummies for mother’s education and father’s education, dummy if any gang in neighborhood, county level variables (Share of white population, black population, Indian population, Hispanic population, share of population under 5 years old, 5-17 years old, 18-20 years old, 21-24 years old, 25-34 years old. 35-44 years old, 45-54 years old. 55-64 years old, 65-74 years old and 75+ years old share of male in population). State tax rates (cigarettes tax and beer tax). There are 85 variables in total, counting each dummy of categorical variable as a variable. 2. White (heteroscedasticity-robust) standard errors are in parenthesis. 3. All of the test statistics are robust against heteroscedasticity. 79 Table 3-4: Incidence of Substance Use using within School Duplication Data (Almost school random & fixed effect estimation.) (1) (2) (3) (4) £5) 16) Dependent van'able MW Wild W W mm mm Method of estimation Random- F ixed- Random- F ixed- Random- Fixed- Effect Efi‘ect Effect Effect Effect Effect mm Peer smoke 0.218 0.218 (0.018) (0.018) Peer drunk 0.310 0.305 (0.019) (0.021 ) Peer drug 0.228 0.221 (0.016) (0.016) 2 ( Other control Yes Yes Yes Yes Yes Yes variables) Hausman Test 1.42 1.20 1.68 F (35, 1201) (0.014) (0.124) (0.000) Single variable 0.680 0.740 -0.625 Hausman test t-statistics Sample size 6312 6312 6312 6312 6312 6312 Note. 1. School id is not available in data. so the (almost) school id is created out of county id, school size, and student -teacher ratio. Surveyor assigns last two variables based on confidential school id number. 2. Same control variables as in Table 3-4 were included. Some of the variables that do not vary within school dropped in fixed effect 3. Heteroscedasticity robust standard errors are in parenthesis for estimated coefficient. For test statistics, p- values are in parenthesis. All the tests are robust against heteroscedasticity. 80 Table 3-5: Incidence of Substance Use using Sibling Data (Household random & fixed effect estimation.) (1) (3) (3) (4) (5) (6) Dependent variable Incidence of QM *ttes W W smoking in last 30 days @nking in last 30 days 111mm Method of Random- F ixed-Effect Random- F ixed—Efi‘ect Random- Fixed-Effect estimation Effect Effect Effect W193) Peer smoke 0.207 0.122 (0.029) (0.040) Peer drunk 0.326 0.247 (0.033) ( 0.044) Peer illegal drug 0.245 0.196 (0.025) (0.034) 2 (Other control Yes Yes Yes Yes Yes Yes variables) E[z Ix] Yes Yes Yes Yes Yes Yes (Contextual Effect) Hausman Test 1.74 1.69 1.21 F (35, 1201) (0.000) (0.000) (0.0%) Single variable -3.830 -2.899 -2.944 Hausman test t-statistics Sample size 2458 2458 2458 2458 2458 2458 Note 1. The same control variables as in Table 3-4 were included. Some of the variables that do not vary within household dropped in fixed effect. 2. Heteroscedasticity robust standard errors are in parenthesis for estimated coefficient. For test statistics. p- values are in parenthesis. All the tests are robust against heteroscedasticity. 81 Table 3—6: OLS Estimates of Incidence of Substance Use (l) (2) (3) Dependent variable mm Insidensmf Incidmsmf ~ oigarottos smoking aloohol drinkigg in W in last ,30 do ys last 30 days lost 30 days OLS OLS OLS Poe; Peer's substance usage (portion) 0.417 0.440 0.352 (0.034) (0.038) (0.030) Female x peer’s usage -0.014 0.050 -0.059 (0.028) (0.033) (0.026) Black x pccr‘s usage -0.3 12 -0.236 -0.1 75 (0.032) (0.038) (0.029) Hispanic x pcer‘s usage -0.l28 -0.060 -0.071 (0.039) (0.044) (0.034) Both biological parents -(). 122 -0.027 -0.047 x peer‘s usage (0.029) (0.033) (0.027) 2 (Other control variables) Yes Yes Yes R2 0143 0.150 0121 Sample size 6356 6356 6356 Note. The same note as Table 3-4 applies. 82 Percentage of users F igure 3—1: Youth Substance Use, Age 12-17, Current Users ——o—— Marijuana and hashish —a—Alcohol —9— Cigarettes 35— 30— 254 20‘ 154 10‘ 5— 1 9190 l l I (Source: National Household Survey on Drug Abuse, annual.) 83 19257