WPS8107 Policy Research Working Paper 8107 Fast, Easy and Cheap Job Matching Social Networks in Bangladesh Norihiko Matsuda Shinsaku Nomura Education Global Practice Group June 2017 Policy Research Working Paper 8107 Abstract This paper uncovers the reason why social networks are used paper concludes that social networks play the role as fast in a job market. The data are novel: a nationally represen- and easy but narrow-spectrum matching. That is, social net- tative matched employer-employee data set in Bangladesh works allow job seekers to find jobs quickly and easily and with detailed information, including direct measures of thereby reduce search costs, but the types of jobs available the use of social networks. The empirical analysis shows from social networks are narrower than those from open that compared with those who used open channels to find channels. As a consequence, those who choose to use social jobs, the employees who used social networks found jobs networks are more likely to end up having mismatched more easily, have lower observable abilities, and achieved jobs, that is jobs in which they cannot take advantage of lower employment outcomes conditional on observable their specialties. In the context of developing countries, a and unobservable abilities. These results are robust whether considerable number of poor job seekers may use social firm-occupation fixed effects are controlled for or not. By networks out of necessity even if the returns to finding good- comparing these findings with theoretical predictions, the match jobs through open channels are sufficiently high. This paper is a product of the Education Global Practice Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted atsnomura@ worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Fast, Easy and Cheap Job Matching: Social Networks in Bangladesh∗ Norihiko Matsuda, University of Wisconsin-Madison † Shinsaku Nomura, World Bank ‡ JEL Codes : J64, J63, J31 Key words : Social networks, referrals, job search, labor markets ∗ The authors are grateful to Yoko Nagashima and Iffath Sharif for their review and feedback. The findings, interpretations, and conclusions in this paper are entirely those of the authors and should not be attributed to the World Bank or its affiliated organizations. † Corresponding author. Department of Agricultural and Applied Economics. Madison, WI. nmatsuda@wisc.edu ‡ Education Global Practice. Washington, D.C. snomura@worldbank.org 1 Introduction There is a consensus in the literature that social networks are widely used in job markets across the world.1 While there are various theories about the roles of social networks in labor markets, mixed empirical evidence does not allow the literature to reach a consensus about the roles. This paper uncovers the role of social networks in a job market, i.e., why social networks are used, by testing theoretical predictions under major hypotheses in the lit- erature. The hypotheses examined are information transmission, screening, peer effects, and nepotism hypotheses. Using nationally representative matched employer-employee data in Bangladesh, we comprehensively estimate empirical counterparts of the theoret- ical predictions based on different regression specifications including a firm fixed effect model and a firm-occupation fixed effect model. We find that employees who used social networks found jobs more quickly and easily, have lower observable abilities and similar unobservable ones, have a higher turnover, earn less salary persistently across job tenure, and are seemingly less productive, com- pared to employees at either the same or different firms who found jobs through formal channels. These empirical results agree with the information transmission hypothesis where job information via social networks and via formal channels originates from dif- ferent sources. On the other hand, the results reject the screening, peer, and nepotism hypotheses. Given the correspondence between the empirical results and the predictions, we con- clude that the role of social networks in the Bangladeshi labor market is fast and easy but narrow-spectrum matching, as suggested by Bentolila et al. (2010) in the case of the U.S. and Europe. That is, social networks help job seekers quickly and easily find jobs, but the types of jobs available from social networks are narrower than those available from open channels. By using open channels, job seekers can find good-match jobs although the search costs are higher since they have to go through longer search spells and apply for more vacancies.2 By using social networks, on the other hand, they can reduce search costs while social networks may bring mismatched jobs. Given this trade-off, those for whom job search through open channels is not worthwhile choose to use social networks and accept mismatched jobs. On the employer-side too, we find evidence that social networks are used as fast 1 For example, Topa (2011); Ioannides and Datcher Loury (2004) for developed countries, and Beaman and Magruder (2012); Larsen et al. (2011); Heath (2016) for developing countries. 2 Throughout this paper, a good-match job means a job in which a worker can exploit his specialty. A mismatched job is a job in which he cannot exploit his specialty. 2 and easy job match. The employers who used social networks as a common mode of job advertisements filled vacancies in a shorter period and felt less difficulty in filling vacancies compared with the employers who did not so commonly use social networks. This finding suggests that social networks allow employers to quickly and easily fill vacancies. Social networks do not necessarily bring the best workers, but employers set lower wages for worse workers. In developing countries, many job seekers may have insufficient financial sources for job search and be very responsive to search costs. In particular, there may be a substantial number of poor job seekers who cannot help but using social networks due to lack of credits even if their returns to finding good-match jobs through open channels are sufficiently high. This paper relates to the broad literature on social networks in labor markets (Bea- man, 2016; Topa, 2011; Ioannides and Datcher Loury, 2004 for excellent reviews). We contribute to it by providing solid empirical evidence showing that social networks serve as fast and easy but narrow-range job matching. This aspect of social networks in la- bor markets, as well as the negative consequence that network-based job search causes mismatches, has been largely missed in the literature. The aspect is studied only by Bentolila et al. (2010) and Pellizzari (2010), to the best of our knowledge. Most of the previous studies highlight other aspects, such as implied by the screening and the peer effects hypotheses, which lead to the positive consequence that social networks generally improve labor market efficiency and do not bring about job mismatches. Our empirical findings that social networks use is associated positively with turnover and negatively with salary are exceptional. The literature mostly finds opposite results. For example, using novel datasets of nine large firms, most of which are based in the U.S. and all of which have employee referrals programs, Burks et al. (2015) show that referred workers have lower turnover and similar or higher earnings. Brown et al. (2016) use data of a single firm and observe that referred workers have lower turnover and higher wage. Simon and Warner (1992) and Kugler (2003) also identify positive associations between social networks use and wages in the U.S. Outside the U.S., Hensvik and Skans (2016) and Dustmann et al. (2016) identify the positive associations with wages in Swe- den and Germany, respectively. Conducting experiments in an online job marketplace, Pallais and Sands (2016) observe that referred workers have lower turnover. By con- trast, only a handful of papers find similar results to ours. Using datasets from the U.S. (the Multi-City Study of Urban Inequality 1992–1994) and Europe (the European Com- munity Household Panel, ECHP), Bentolila et al. (2010) find that workers who found jobs through social networks earned less. Pellizzari (2010) uses the same European data 3 (the ECHP) as Bentolila et al. (2010) and identifies the negative association between social networks use and wages in some countries. In the case of Ethiopia’s cut flower farms, Mano et al. (2011) find that workers recruited through social networks received discounted wages at entry, which quickly disappeared with job tenure, however. This paper also adds to the literature by providing clean and country-level evidence. First, our estimation results are robust to controlling for firm fixed effects and firm- occupation fixed effects. In other words, the associations of social networks use with salary, turnover, and other outcomes do not change no matter whether the referred work- ers are compared with the nonreferred at the same firms or at different firms. Controlling for firm fixed effects is possible only with matched employer-employee data. Second, we use direct measures of social networks use, which are based on responses of employees. Besides, we check the validity of the measures with employer-side information. We do not need to rely on indirect measures such as being neighbors (e.g., Schmutte, 2015) and having worked at the same firms previously (e.g., Hensvik and Skans, 2016). Lastly, our data are nationally representative while most of the empirical evidence in the literature builds on data from small geographical areas or specific firms. 2 Theories in the literature The literature has various theoretical models of social networks in labor markets (Bea- man, 2016; Topa, 2011; Ioannides and Datcher Loury, 2004 for reviews). Although those models hypothesize different roles of social networks, or different reasons behind social networks use, most of the hypotheses can be categorized into the following: transmission, screening, peer effects, and nepotism hypotheses. We briefly explain each hypothesis and related literature. We summarize the theoretical predictions under each hypothesis in table 1. First, the transmitting information hypothesis is that social networks play the role in transmitting more information about job vacancies. Building on this hypothesis, a large number of theoretical papers study labor market consequences of social networks use (e.g., Calv´o-Armengol and Jackson, 2004, 2007; Ioannides and Soetevent, 2006). While most of these papers are dedicated to analyzing network-level outcomes such as neighborhood and ethnicity rather than worker-level analysis, they have individual-level predictions that referred workers experienced a shorter unemployment spell and earn more than nonreferred ones. The intuitions behind the predictions are that since re- ferred workers are connected to more contacts on average and therefore receive more job 4 information, they find jobs more quickly and have higher reservation wage.3 The predic- tions are summarized in column (1) of table 1. Empirically, Munshi (2003) and Beaman (2012) find supporting evidence. Importantly, the predictions require the assumption that the job information via social networks and via open channels originates from an identical underlying distribution of job vacancies. This assumption is common among the studies building on the transmitting information hypothesis. Notably, Bentolila et al. (2010) make the distinct assumption that job information sources are different while they still view the primary role of social networks as trans- mitting vacancy information. To our knowledge, their paper is the only one having this assumption. The settings are as follows: there are two types of occupations; each worker has occupation-specific advantage in a randomly chosen occupation;4 this advantage is observable to everyone; her social network is connected to one occupation, and she can find more easily a job at the connected occupation than a job at the unconnected occu- pation. If her specialized occupation agrees with the connected occupation, she searches for a job of her specialty. If they disagree, she may not bother searching for a job in her specialty but rather take up a job that she can find through the network. As a result, those who use networks are more likely to work for mismatched jobs, and hence they earn less on average. Column (2) in table 1 summarizes the predictions. Noteworthy is the negative association between wages and the use of social networks, which is unique to their models. Although their model is simplified to having two occupations, in our understanding, the key difference from the other models is that information via social networks is drawn from a different distribution than the one from which information via open channels is drawn. Thus, we call their assumptions about the role of social networks the different information transmission hypothesis. Second, the screening hypothesis is that social networks reduce uncertainty about a worker’s productivity or match quality by bringing hard-to-observe information. In the seminal paper of Montgomery (1991), workers are observationally equivalent, and the productivity of workers is ex ante unknown and revealed after employment.5 Wages cannot be contingent on output, and workers live only for one period. As a role of so- cial networks, the model assumes inbreeding, or homophily, of social structure: workers within a network are relatively homogeneous. Thus, firms are able to more precisely 3 o-Armengol (2004) and Calv´ Calv´ o-Armengol and Zenou (2005), however, show non-monotonicity of unemployment probability with respect to network size. 4 This occupation specific advantage can be interpreted in various ways. It may represent occupation- specific expertise, personal preference which leads to high future productivity, or any other types of match quality. 5 The model holds whether workers know their productivity ex ante or not. Thus, it is possible to consider the productivity as match quality. 5 infer the productivity of applicants who belong to networks of incumbent workers. The model predicts that referred workers earn more and have higher unobservable abilities than nonreferred workers conditional on observable abilities.6 Hensvik and Skans (2016) provide empirical evidence supporting the predictions. Although the model of Mont- gomery (1991) does not have a prediction about wage growth since a worker lives for one period, we straightforwardly expect that wage growth is higher for referred workers since they have better abilities on average. Another seminal paper is Simon and Warner (1992). Although their framework, ap- plying Jovanovic’s (1979) matching model, is different from Montgomery (1991), they essentially assume the same role of social networks: social networks reduce the ex ante uncertainty about match quality. Since their model is dynamic, they have predictions about turnover and wage growth, which Montgomery (1991) does not have. The set- tings are the following: Infinitely-lived workers meet employers through either social networks or open markets; workers are homogeneous in the sense that their firm-specific productivity, or job-match quality, is independently and identically distributed; the pro- ductivity is ex post observable but is ex ante only partially observable with noise to both workers and employers; wages can be renegotiated, and workers can quit jobs; lastly, the role of social networks is modeled as that the noise is smaller when workers and employers meet through social networks than when they meet in open markets. In an equilibrium, entry wage is set equal to the expected productivity, and subsequent wage is equal to the actual productivity. The predictions are the following: (i) Entry wages are higher for those who find jobs through social networks than those who find them through open channels. (ii) Wage growth of nonreferred workers is higher. (iii) Referred workers have lower turnover than nonreferred ones.7 , 8 Brown et al. (2016) and Dustmann et al. (2016) obtain empirical results aligning with these predictions. The predictions under 6 The original paper does not explicitly state the prediction about unobservable abilities, however. 7 The reason for prediction (i) is that the nonreferred, whose productivity is less revealed at hire, have the option to quit jobs if their firm-specific productivity turns out low while they can enjoy high wage if the productivity turns out high. The reason for prediction (ii) is the following: wages from the subsequent period are set exactly equal to the true productivity, and only those whose productivity is higher than an identical threshold irrespective of being referred or not continue to work. The reason for prediction (iii) is that the probability that actual productivity turns out higher than the threshold conditional on an initial offer being accepted is higher for the referred than the nonreferred. 8 These predictions hold in extended versions such as Dustmann et al. (2016) and Galenianos (2013). Dustmann et al. (2016) slightly formalize the original model and confirm that the same predictions are derived. Galenianos (2013) adds heterogeneity in firm’s productivity and allows firm’s investment in reducing the ex ante uncertainty particularly when hiring through an open labor market. Even in this case, the same predictions about entry wages, wage growth, and turnover are obtained conditional on firm’s productivity. However, unconditional association between social networks use and wage can be negative. 6 the screening hypothesis are shown in column (3) of table 1. Third, the peer effects hypothesis is that social networks are exploited to increase worker’s performance by peer effects. Kugler (2003) puts forth the model where moral hazard occurs. She models the role of social networks as that those who are referred by high-effort workers have disutility from shirking due to peer monitoring by their referrals. Put intuitively, peer monitoring lowers employer’s costs for monitoring.9 Predictions are that referred workers earn more and have longer job tenure than nonreferred ones.10 Different aspects of peer interactions, other than peer monitoring, are studied in other papers empirically. For example, Bandiera et al. (2013) examine how team productivity differs depending on incentive structures; Bandiera et al. (2005) examine how individuals respond to potential externalities to their peers; Pallais and Sands (2016) experimentally test whether referred workers are more productive when working with referrers and whether referred workers exert more effort when they know that their performance known to referrers affects referrers’ promotion. They do not have theoretical predictions relevant to our empirical analysis, however.11 The predictions are summarized in column (4) of table 1. Lastly, the nepotism hypothesis is that nepotism by incumbent workers and employ- ers is the reason behind social networks use. Nepotism plausibly takes place in job markets if private returns of incumbent workers and employers to referral differs from firm’s returns to it. A few papers empirically find the existence of nepotism (Wang, 2013; Fafchamps and Moradi, 2015), but, to our knowledge, no theoretical models con- ceptualize the use of social networks as nepotism.12 Nonetheless, we expect that under 9 Another assumption on social networks in her model is that social networks are less efficient than the open market in terms of the encounter rates. While this assumption may not be plausible in some contexts, it is mainly for the convenience in having the two matching channels exist in an equilibrium. It does not drive main implications. 10 The reason for the wage difference is that firms pay efficiency wages to referred workers while firms pay lower wages to nonreferred workers and incur shirking. The reason for the tenure difference is that non-referred workers end up being dismissed due to their shirking. 11 Heath (2016) develops a model which incorporates moral hazard problems and produces predictions about observable abilities and wage growth. Her model is quite specific to the empirical context where a minimum wage regulation is binding. A firm may be reluctant to hire a low-ability worker if willingness- to-pay wage is below the minimum wage. In this case, if a low-ability worker has a high-ability worker in his network, he and the firm can get around the minimum wage restriction by bundling up the wages of him and the high-ability companion. This bundling of wages is the role of social networks in her model, and there are no peer effects. Thus, we do not categorize the model under the peer effects hypothesis although it is sometimes categorized so (e.g. Beaman, 2016). The model is also called as a moral hazard model and viewed as the same category as Kugler (2003) (e.g., Brown et al., 2016). However, we differentiate it from Kugler (2003) since the bundling to get around the minimum wage, which is the key role of social networks in the model, does not require moral hazard and applies to the case where worker’s output is perfectly observable. 12 Goldberg (1982) puts forth a theory modeling nepotism. However, it focuses on firm-side implica- 7 the nepotism hypothesis, referred workers experienced a shorter job search spell, have lower abilities and productivity.13 We consider that implications on turnover and wages are theoretically unclear and depend on contexts. Column (5) in table 1 summarizes the predictions. 3 Data We use Bangladeshi matched employer–employee survey data collected in 2012 by the World Bank. The data are nationally representative of formal-sector employers and employees in manufacturing, commerce, finance, education, and public administration.14 The survey randomly sampled establishments, and then sampled employees within the establishments. The sampling of establishments was stratified by industries and size.15 Interviews were administered separately for employers and employees. Note that while the data represent the formal sector employees, they do not represent the whole labor force since the data do not include informal-sector employees or the unemployed labor force. We restrict the sample as follows. We include only three industries: manufactur- ing, commerce, and finance, because the other industries: education and public admin- istration, consist mostly of government or government-aided organizations for public services.16 The three industries account for 48% and 71% of the establishments and employees, respectively, in the Bangladesh formal sector (Nomura et al., 2013). We ex- clude seasonal and day laborers and part-time and contract workers, and focus only on full-time permanent employees,17 who constitute about 90% of the original sample. We also restrict the sample to males, who account for nearly 90% of the original sample. Lastly, we restrict the sample to those aged 40 and below, who constitute about 90% of the original sample. The reason for excluding older employees is that information at the timing of recruitment, such as initial salary, is crucial for our analysis and may be less accurate for older employees since they tend to have been recruited a longer time ago. After the above restrictions are applied, our sample has 339 establishments and 3,396 tions of nepotism and does not have predictions relevant to our study. 13 In a Bangladeshi context, Mahmood and Absar (2015) suggests that there exists nepotism favoring relatives and friends in a labor market. 14 See Nomura et al. (2013) for details about the data. 15 The size is defined by the number of employees and consists of small (10–20), medium (21–70) and large (71+). 16 Most schools in Bangladesh are government-aided privately-managed schools. The recruitment of full-time teachers is regulated by the government. 17 Permanent workers are those who have an indefinite period of employment. 8 employees. The data includes information about cognitive and non-cognitive skills that were assessed during the survey based on short tests and questions about language and math skills and personality traits. These skills are likely to be unobservable, or not easily observable, to employers at the timing of hiring. The non-cognitive skills include Big 5 personality traits, which we convert into a single index for our analysis.18 Summary statistics of establishments are shown in table 2. Approximately, half the sample is manufacturing (N = 191), and the other half is evenly split into commerce and finance (N = 74 each). The proportions of small, medium and large establishments are respectively 48%, 29%, and 22%. 61% are single establishment firms. 69% use personal networks as a common mode of job advertisement. The ways employees found their current jobs are presented in table 3. Social networks are the most prevalent channel to find jobs. 63% of employees found jobs through social networks. As for social network type, families and relatives, friends, and neighbors were dominant, used by 23%, 27%, and 12%, respectively. Political affiliations and school alumni were almost negligible, used by only 1 % each. Among formal and open channels, the most important is media advertisements, through which 33% of employees found jobs. The other formal methods such as public and private employment services and internet postings were used only by 4%. These statistics demonstrate that social networks of families/relatives, friends, and neighbors are the most common job search method in the formal sector and that media advertisements are dominant among the formal open methods. Our measure of social networks use is based on employees’ responses as to whether they found a job through social networks or not. This measure, however, does not say whether an employee was referred or not. The data do not include information about referral. Thus, the use of social networks in this paper is not restricted to referrals but includes various forms such as passing on information. Table 4 shows summary statistics of employees. An average employee is 24 years old and has 9.0 years of schooling and 4.3 years of tenure. The composition of entry occupations is that 6%, 20%, 61% and 14% of employees were recruited as managers, professionals, semi-skilled workers, and elementary workers, respectively. Only 4% have been promoted in current firms. Compared to those who did not use social networks to find jobs, those who did are younger and less educated, have slightly shorter tenure 18 For Big 5, we firstly calculate the z-score for each of the five personality traits, i.e., openness, consci- entiousness, extroversion, agreeableness, and emotional stability, using the sample mean and standard deviation. Then we construct Big 5 index as an equally weighted average of them, as Burks et al. (2015) do. 9 and less earnings, are more likely to be at elementary occupations and less to be at professional occupations, and less often experienced promotions. 4 Estimation strategy Our overall strategy to uncover the role of social networks is to estimate differences in job search experience, abilities, and employment outcomes between referred and nonre- ferred workers and then compare empirical results to theoretical predictions. We run the following regression equation: yijk = θN etworkijk + β Xijk + γjk + ijk , (1) where i, j , and k denote employees, establishments, and entry occupation types, respec- tively; N etworkijk is a dummy indicating that employee i found his job through social networks; Xijk are controls. The coefficient θ represents a difference associated with the use of a social network. It is important to note that our estimate of θ, a difference associated with the use of social networks, is essentially descriptive, not causal. Therefore, if there is an unobserv- able factor that is associated with the use of social networks and a dependent variable and that is not endogenized in theoretical models, our estimate is unable to test theo- retical predictions. To address this endogeneity problem, we have different specifications of the fixed effect, γjk , and the controls, Xijk , and draw conclusions that are consistent across the different specifications.19 The three different specifications of the fixed effect, γjk control for no fixed effects, firm fixed effects, and firm-occupation fixed effects, respectively. Therefore, the first specifi- cation estimates an association of social networks usage in the whole sample across firms, whereas the second and third specifications estimate an association, respectively, within firms and within firm-occupation cells. The third specification is not necessarily more preferable than the others. For example, if the use of social networks causes sorting of employees into different firms, estimates from the first specification are meaningful. But, if employee’s selection into firms is driven by an unobservable factor, the second and third specifications are more informative.20 Hence, we consider that the three specifications 19 It is worth noting that most of the previous studies rely on descriptive analyses. Exceptional papers exploiting exogenous variation in social networks use are Beaman (2012) and Pallais and Sands (2016). Besides, among the descriptive studies, only a handful of papers, such as Dustmann et al. (2016), provide estimates with and without controls for firm fixed effects. 20 Since the theoretical predictions, except for those under the information transmission hypothesis, apply to within-firm differences in outcome variables, within-firm comparison based on the second and 10 are all informative and complement each other. A concern in our estimations is that they may suffer from attrition bias since our data do not include those who had been employed in the formal sectors but quit already. For example, if the use of social contacts is associated with turnover patterns, our estimates may have selection bias. We examine this attrition bias issue later. 5 Results 5.1 Job search spell Table 5 presents differences in terms of last job search experience between employees who used social networks and those who did not. Those who used social networks found jobs in a significantly shorter spell. The difference not controlling for firm fixed effects is −24% (column (2)), and the within-firm and the within-firm-occupation difference are −20% and −22%, respectively (columns (3) and (4)). Besides, those who used social networks applied for fewer vacancies by 16–23% (columns (6)–(8)), which implies that a shorter search spell was not a result of more intensive search. These results show that those who used social networks found jobs more quickly and easily. 5.2 Observable abilities Table 6 presents differences in the observable abilities: years of schooling, age at hire, and parent’s education. The estimation results across the different specifications clearly show that those who used social networks have lower observable abilities compared with those who used formal channels. According to the estimates controlling for firm-occupation fixed effects, those who used social networks have significantly less schooling by 1.39 years (column (3)) and were younger by 0.61 year at hire (column (6)). Their fathers and mothers are less likely to have completed primary school education by 7% and 9% points, respectively (columns (9) and (12)).21 third specifications is informative. The predictions under the same information transmission hypothesis may or may not be valid for within-firm differences, though. 21 The result that those using social networks have less schooling could have been driven by the possibility that job seekers with poor family backgrounds tend to rely on networks. A regression of years of schooling with parents’ education being included as a control for family backgrounds, however, gives a very similar coefficient of the social network dummy in terms of magnitude and significance, which implies that the use of social networks is negatively associated with schooling conditional on similar family backgrounds. 11 5.3 Unobservable abilities Table 7 presents estimates of differences in seemingly unobservable cognitive and non- cognitive abilities: the math, language, and personality scores. While these abilities were made observable to econometricians by conducting an assessment in the survey, they are somewhat unobservable, or hard-to-observe, to employers during a recruitment process. The associations between social networks and the math and language scores are statistically insignificant conditional on the observable abilities, except for the regression of Big 5 without fixed effects (column (10)). The associations may also be economically insignificant. According to the estimations in columns (4) and (8), those who used social networks have higher math and language scores at most by 0.19σ and 0.24σ , respectively, based on the upper bound of the 95% confidence interval.22 As for the Big 5 index, the within-firm-occupation difference is precisely estimated to be zero (column (12)) while the difference not controlling for firm-fixed effects may be economically significant. We interpret the above results as indicating that the workers who used networks and those who did not are precisely similar in terms of the unobserved abilities, especially within colleagues at the same occupations in the same firms. 5.4 Tenure The average job tenure of those who found jobs through social networks is shorter by 12– 20% (columns (2)–(4) in table 8). These estimates are significant at the 10% level, except for that controlling for firm fixed effects (column (3)). It is important to understand that this estimation compares the average tenure of current employees who had survived, not the expected tenure conditional on being hired. Thus, the estimation results do not mean that the expected tenure conditional on being hired is 19% shorter for the employees who used social networks than for those who did not use them. Nonetheless, the results are likely to suggest that the use of social networks is negatively associated with job tenure. 5.5 Entry salary and salary growth The regressions of entry salary, including bonuses, overtime, and other compensations, reported in table 9 show that those who used social networks received significantly lower entry salary by 7–9% conditional on the observable and unobservable abilities whether 22 In the literature on impact evaluation of certain interventions, an effect size of about 0.2 on edu- cational outcomes such as cognitive scores is considerable. However, in our context where descriptive associations, not impacts of a treatment, are examined, a 0.2σ difference in cognitive skills is not so considerable to the whole distribution of the skills. 12 firm and firm-occupation fixed effects are included or not (columns (2)–(4)).23 To examine salary growth, we run the following equation: log(CurrentW age)ijk = θN etworkijk +ρN etworkijk × tenureijk + β Xijk + γjk + ijk . (2) Job tenure as well as its interactions with individual characteristics such as abilities are included in Xijk . The coefficient of the interaction between the network dummy and tenure, ρ, represents a salary growth difference between those who used social networks and who did not. The coefficient of the network dummy, θ, represents a difference in entry salary. The estimation results are reported in table 10. First, the coefficient of the network dummy confirms that those who used social networks received significantly lower entry salary. The entry salary discount is 9–13% (columns (2)–(4)), which are similar to the estimates, 7–9%, reported in table 9. This similarity indicates robustness of the esti- mated entry salary discount. Second, the difference in annual salary growth is precisely estimated to be very small, if any, which implies that the salary discount does not dissi- pate with job tenure. According to the estimation controlling for firm-occupation fixed effects (column (4) in table 10), the salary discount initially starts at 11.8% and narrows only by 0.5% annually according to the point estimates. The salary gap persists at 9.3% after five years and at 6.8% after ten years.24 In addition, this annual catch-up is much smaller than the annual growth, 6.7%, which is common to the both types of employees. We conduct a different investigation on salary growth. Specifically, we run equation (1) with the dependent variable being a current salary separately for a subsample of each tenure, and examine how the salary gap changes with tenure. The motivation for this analysis is to address the possibility that a salary trajectory may not be linear. The result is illustrated in figure 1. It shows that the salary gap is persistent at about 10% until 7 years of tenure. This persistent gap corroborates the finding that the earnings of those who used social networks do not catch up. 23 The coefficient of the language score is negative. This is opposite our expectation, and we do not have a good explanation for it. 24 According to a formal statistical test that takes into account the variances and covariance of the two coefficients, the hypothesis that the salary gap is zero after 10 years is rejected with the p-value being 0.0473. The corresponding p-value for 11 years is 0.0996. 13 5.6 Productivity: Promotion Our data do not have a direct measure of employee’s productivity. We use promotion history as a proxy of employee’s productivity, following Brown et al. (2016).25 Regres- sion results where the dependent variable is the dummy for having ever been promoted is reported in table 11. The promotion likelihood for those who used social networks is about 1–2% points lower in all the specifications although the estimates are not signifi- cant.26 This result may suggest that those who used social networks are less productive although the result lacks statistical precision and also may suffer from measurement errors of productivity. 5.7 Robustness checks 5.7.1 Attrition bias A concern in our results is that the estimates may have attrition bias. This is because our data do not include former employees who had already quit before the data collection and because our estimation on tenure suggests that turnover is correlated with the use of social networks. To examine if the attrition bias exists, we run the same estimations using the subsample of those who were hired in the last three years. Although the attrition may have occurred among those new workers, the subsample may suffer less from the potential attrition bias. Appendix A presents all the estimation results using the subsample. Although those who used social networks still have shorter tenure by 9% than those who did not (column (4), table A1), we obtain basically the same results. We examine the attrition bias issue in a different way. We estimate a difference between those who used networks and who did not, separately using a subsample of each tenure. If the attrition bias is in play, the estimate may differ with tenure. Since estimates for shorter tenure are less affected by the attrition, we particularly look into if estimates for shorter tenure are similar to the main results explained in the preceding subsections. Figure A1 demonstrates the estimation results. Although they lack pre- cision, the estimates, particularly the ones for short tenure, look similar to the main results. 25 Although promotion plausibly correlates with productivity, the obvious limitation is that promotion reflects only a subset of productivity and also correlates with different factors than productivity. 26 Since the sample mean of the likelihood is 4% (table 4), this 1–2% points difference is substantial. 14 5.7.2 Other robustness checks We additionally conduct two robustness checks. First, since we restrict the sample to the employees aged 40 and below, we examine if our results are sensitive to the cutoff age by running the same estimations for the sample aged 60 and below. We obtain similar estimation results (appendix B). Second, our main estimations use the single aggregated index of Big 5 rather than the five indexes of personality traits. As a robustness check, we run the same estimations using the five indexes individually and find similar results (Appendix C).27 6 Discussions The summary of the results, reported in column (6) of table 1, is that employees who used social networks found jobs more quickly and easily, have lower observable abili- ties and similar unobservable ones, have a higher turnover, earn less salary persistently through job tenure, and are seemingly less productive, compared to those who found jobs through formal channels. These results are robust whether firm fixed effects and firm-occupation fixed effects are controlled for or not. This robustness somewhat implies that our estimation results are not driven by unknown factors that are not incorporated in the hypothesized theories. In the following subsections, we first compare these results with the theoretical pre- dictions under the different hypotheses. Then, we discuss the role of social networks in the Bangladeshi labor market. 6.1 How the empirical findings agree with the theoretical pre- dictions The same information transmission hypothesis disagrees with our results. Under the hypothesis, social networks increase job arrival rate and thereby reservation wage. Hence, wages of referred workers are predicted to be higher on average. This prediction is robust across different models.28 Contrary to the prediction, however, our estimations strongly show that the use of social networks is negatively associated with wages. 27 From a different perspective other than the robustness check, these estimations are interesting in that the estimations may inform us how each trait is associated with labor market outcomes. 28 In the case that members in social networks increase, social networks can be negatively associated with wages in a short run after the increase. However, this temporary negative association with wage is not relevant to our results since there is no reason to believe that such increase in network members occurred across Bangladesh. Long-run implications correspond to our empirical context. 15 The screening hypothesis also contradicts our results. Under the hypothesis, since social networks bring hard-to-observe information and reduce uncertainty about pro- ductivity, referred workers are strongly predicted to be more productive, earn more, and have lower turnover, compared to nonreferred workers particularly within the same firms. However, our results controlling for firm-fixed effects reject the predictions on wage and turnover. The peer effects hypothesis does not match the results either. Since social networks are exploited to increase workers’ productivity or extract workers’ effort, referred workers are more productive and earn more. They also have lower turnover since they do not shirk and therefore are not dismissed.29 The estimation results imply the opposite, however. As for the nepotism hypothesis, although the discrepancy between the predictions and the results is not strong, we interpret that the empirical results do not support it for two reasons. First, the empirical finding that referred workers earned about 10% less significantly than nonreferred ones conditional on abilities seems not to support that nepotism is in play although nepotism does not necessarily lead to patronees earning more than they would earn in the absence of nepotism. We consider that the 10% salary discount is too economically significant for nepotism to exist behind the use of social networks.30 Second, if nepotism is in play, referred workers may have higher likelihood of promotion than nonreferred ones since nepotism may help promotion. However, the empirical analysis suggests referred workers faced lower likelihood of promotion than nonreferred ones conditional on abilities although this difference is not significant. Lastly, the different information transmission hypothesis agrees with the empirical results well. Under the hypothesis, job search through social networks causes mismatch because “workers may sacrifice their productive advantage so as to find a job more easily” (Bentolila et al., 2010). Hence, referred workers are less productive, earn less, and have higher turnover. All of these predictions are supported by the empirical findings although the empirical result on productivity is not significant. The employer-side information also supports the different information transmission hypothesis. According to table 12, the employers who use social networks as a common mode of job advertisements fill vacancies in a shorter period than those who do not (2.6 weeks vs 4.8 weeks). Also, the employers using social networks feel less difficulty in filling job vacancies. These findings suggest that social networks help employers find 29 These predictions may be more relevant for within-firm estimates. 30 A literature on patronage in Bangladesh suggests that favoritism for relatives and friends is salient at workplaces (e.g., Mahmood and Absar (2015)). Such favoritism may help patronees get not only jobs but also high salary and promotion. However, this is not supported by our empirical findings, especially the economically significant wage discount conditional on abilities. 16 employees quickly and easily. The findings seem to agree with the information trans- mission hypothesis. On the other hand, the findings do not necessarily agree with the screening and the peer effect hypotheses since under these two hypotheses the employers using social networks may spend longer time on recruitment in order to carefully select applicants.31 6.2 The role of social networks Given the correspondence between the empirical results and the theoretical predictions, we conclude that the role of social networks in the Bangladeshi labor market is fast and easy but narrow-spectrum matching. That is, social networks help job seekers quickly and easily find jobs, but the types of jobs available from social networks are narrower than those available from open channels. By using open channels, they can find good- match jobs although the search costs are higher since they have to go through longer search spells and apply for more vacancies. By using social networks, they can reduce search costs while they may find only mismatched jobs. Given this trade-off, those for whom the extensive but expensive job search through open channels is not worthwhile choose to use social networks and end up having mismatched jobs. As a consequence, the jobs filled by social networks are more likely to be mismatched than those filled by open channels. Table 13 provides additional evidence consistent with the above interpretation. The table shows main reasons why employees chose the jobs. Compared to those who did not use social networks, those who did are more likely to have chosen the jobs because of “no other offers” and “others’ recommendation” and less likely to have chosen because of “good prospect for career progression” and “good prestige of firm.” This result indicates that those who used social networks sacrifice better match (less “good prospect for career progression”) and have lower abilities like motivation and forthcoming labor market performance (less “good prospect for career progression,” less “good prestige of firm,” and more “no other offers”).32 While Bentolila et al.’s (2010) model does not have predictions about observable abil- 31 Table 12 provides another important implication. Since the social networks dummy we use is based on information from employees, our main results do not directly inform why employers use social networks. It is even possible that employers do not know whom of their employees used social networks. However, the employers who commonly use social networks have a significantly larger proportion of employees who found jobs through social networks: the proportion of such employees is 82.9% for the employers using social networks but 37.0% for the employers not using them. This result implies that employers actually use social networks and know whom of employees use social networks. 32 These abilities seem not to be well captured by years of schooling, or the cognitive or non-cognitive skills. 17 ities, the negative association between social networks and observable abilities identified in our data stands with the above interpretation. It is plausible that the wage difference between a good-match job and a mismatch job increases with abilities. If so, the returns to searching jobs through open channels are higher for high-ability workers, and hence the use of social networks negatively correlates with observable abilities. Another pos- sibility is that if wealth and abilities are positively correlated, lower-ability workers are more likely to be poor, lack credits, and be unable to afford an open-channel search. Importantly, in the context of developing countries where a considerable number of people live on the margin of subsistence, a small difference in search costs may matter a lot, and more job seekers may use social networks to reduce search costs than in developed countries. Besides, if poor job seekers lack access to credits, they simply cannot use open channels but social networks even if the returns to finding good-match jobs through open channels is sufficiently high. Table 14 somewhat suggests that poorer workers are more likely to use social networks to find jobs. The table shows correlations of social networks with worker’s characteristics in a slightly different way than tables 6 and 7. The negative correlations with parents’ education in table 14 may imply that poorer workers tend to use social networks since parents’ education somewhat represents worker’s wealth while it also correlates with worker’s ability. 7 Conclusions Social networks are widely used in job markets across the world. Yet the reason for the widespread use remains unclear. Using the novel data in Bangladesh, this paper provides new evidence about the role of social networks in labor markets. The empirical results find that the employees who used social networks found jobs more quickly and easily, have lower observable abilities and similar unobservable ones, have a higher turnover, earn less salary persistently across job tenure, and are seemingly less productive, compared to the employees, at either the same or different firms, who found jobs through formal channels. Building on these findings, we draw the conclusion that social networks play the role as fast and easy but narrow-spectrum matching. That is, social networks help job seekers easily and quickly find jobs, but the types of jobs offered through social networks are narrower. Social networks, as fast and easy but narrow-spectrum job matching, cause mis- matches. If contacts of job seekers are unluckily at different occupations than their specialties, they face a trade-off between job match quality and search costs, and some 18 of them opt to use social networks and accept mismatched jobs. In the context of de- veloping countries, a reduction in job search costs, primarily in the form of a reduction in search spells, may matter more than in developed countries. Besides, there may be a substantial number of job seekers who cannot help but using social networks because of limited access to credits even if an increase in earnings by finding a good-match job through open channels exceeds the reduction in search costs by using social networks. 19 References Bandiera, Oriana, Iwan Barankay, and Imran Rasul (2005) ‘Social preferences and the response to incentives: Evidence from personnel data.’ Quarterly Journal of Economics 120(3), 917–962 (2013) ‘Team incentives: Evidence from a firm level experiment.’ Journal of the Eu- ropean Economic Association 11(5), 1079–1114 Beaman, Lori (2016) ‘Social networks and the labor market.’ In ‘The Oxford Handbook of the Economics of Networks’ Beaman, Lori A. (2012) ‘Social networks and the dynamics of labour market outcomes: Evidence from refugees resettled in the U.S.’ Review of Economic Studies 79(1), 128– 161 Beaman, Lori, and Jeremy Magruder (2012) ‘Who gets the job referral? Evidence from a social networks experiment.’ American Economic Review 102(7), 3574–3593 Bentolila, Samuel, Claudio Michelacci, and Javier Suarez (2010) ‘Social contacts and occupational choice.’ Economica 77, 20–45 Brown, Meta, Elizabeth Setren, and Giorgio Topa (2016) ‘Do informal referrals lead to better matches? Evidence from a firm’s employee referral system.’ Journal of Labor Economics 34(1), 161–209 Burks, Stephen V, Bo Cowgill, Mitchell Hoffman, and Michael Housman (2015) ‘The value of hiring through employee referrals.’ Quarterly Journal of Economics pp. 1–36 o-Armengol, Antoni (2004) ‘Job contact networks.’ Journal of Economic Theory Calv´ 115(1), 191–206 o-Armengol, Antoni, and Matthew O. Jackson (2004) ‘The effects of social networks Calv´ on employment and inequality.’ American Economic Review 94(3), 426–454 o-Armengol, Antoni, and Matthew O Jackson (2007) ‘Networks in labor mar- Calv´ kets: Wage and employment dynamics and inequality.’ Journal of Economic Theory 132(1), 27–46 o-Armengol, Antoni, and Yves Zenou (2005) ‘Job matching, social network and Calv´ word-of-mouth communication.’ Journal of Urban Economics 57(3), 500–522 Dustmann, Christian, Albrecht Glitz, Uta Sch¨ onberg, and Herbert Br¨ ucker (2016) ‘Referral-based job search networks.’ Review of Economic Studies 83(2), 514–546 Fafchamps, Marcel, and Alexander Moradi (2015) ‘Referral and job performance: Ev- idence from the Ghana colonial army.’ Economic Development & Cultural Change 63(4), 715–751 20 Galenianos, Manolis (2013) ‘Learning about match quality and the use of referrals.’ Review of Economic Dynamics 16(4), 668–690 Goldberg, Matthew S. (1982) ‘Discrimination, nepotism, and long-run wage differentials.’ Quarterly Journal of Economics 97(2), 307–319 Heath, Rachel (2016) ‘Why do firms hire using referrals ? Evidence from Bangladeshi garment factories.’ Unpublished manuscript om Skans (2016) ‘Social networks, employee selection, Hensvik, Lena, and Oskar Nordstr¨ and labor market outcomes.’ Journal of Labor Economics 34(4), 825–867 Ioannides, Yannis M., and Adriaan R. Soetevent (2006) ‘Wages and employment in a random social network with arbitrary degree distribution.’ American Economic Review 96(2), 270–274 Ioannides, Yannis M, and Linda Datcher Loury (2004) ‘Job information networks, neigh- borhood effects, and inequality.’ Journal of Economic Literature 42(4), 1056–1093 Jovanovic, Boyan (1979) ‘Job matching and the theory of turnover.’ Journal of Political Economy 87(5), 972–990 Kugler, Adriana D (2003) ‘Employee referrals and efficiency wages.’ Labour Economics 10(5), 531–556 Larsen, Anna Folke, John Rand, and Nina Torm (2011) ‘Do recruitment ties affect wages? An analysis using matched employer-employee data from Vietnam.’ Review of Development Economics 15(3), 541–555 Mahmood, Monowar, and Mir Mohammed Nurul Absar (2015) ‘Human resource in Bangladesh : Current scenario and future.’ South Asian Journal of Human Resources Management 2(2), 171–188 Mano, Yukichi, Takashi Yamano, and A Y A Suzuki (2011) ‘Local and personal networks in employment and the development of labor markets : Evidence from the cut flower industry in Ethiopia.’ World Development 39(10), 1760–1770 Montgomery, James D (1991) ‘Social networks and labor-market outcomes: Toward an economic analysis.’ American Economic Review 81(5), 1408–1418 Munshi, Kaivan (2003) ‘Networks in the modern economy: Mexican migrants in the U. S. labor market.’ Quarterly Journal of Economics (May), 549–599 Nomura, S, S Y Hong, Christophe J. Nordman, L R Sarr, and A Y Vawda (2013) An assessment of skills in the formal sector labor market in Bangladesh: A technical report on the enterprise-based skills survey 2012 (South Asia Development Sector Report No. 64) (Washington, DC: The World Bank) Pallais, Amanda, and Emily Glassberg Sands (2016) ‘Why the referential treatment? Ev- 21 idence from field experiments on referrals.’ Journal of Political Economy 124(6), 1793– 1828 Pellizzari, Michele (2010) ‘Do friends and relatives really help in getting a good job?’ Industrial and Labor Relations Review 63(3), 494–510 Schmutte, Ian M (2015) ‘Job referral networks and the determination of earnings in local labor markets.’ Journal of Labor Economics 33(1), 1–32 Simon, Curtis J., and John T. Warner (1992) ‘Matchmaker, matchmaker: The effect of old boy networks on job match quality, earnings, and tenure.’ Journal of Labor Economics 10(3), 306–330 Topa, Giorgio (2011) ‘Labor markets and referrals.’ Handbook of Social Economics 1, 1193–1221 Wang, Shing-Yi (2013) ‘Marriage networks, nepotism, and labor market outcomes in China.’ American Economic Journal: Applied Economics 5(3), 91–112 22 .6 Coefficient of the social network dummy -.4 -.2 0 -.6 .2 .4 1 2 3 4 5 6 7 8 9 Tenure in years Figure 1: Differences in current wages between those who used social networks and those who did not Note. Plotted are estimates of θ, the coefficient of the social network dummy, in the equation of log(CurrentW age)ijk = θN etworkijk + β Xijk + γjk + ijk , where Xijk are years of schooling, age at hire, math and language z-scores, Big 5 index, and father’s and mother’s years of schooling, and γjk is an establishment-entry occupation fixed effect. Occupations are grouped into manager, professional, semi- skilled, and elementary. The regressions are run separately for a subsample of each year of tenure, except for 1-year and 9-year tenure. The subsamples of 0- and 1-year tenure are jointly used because their size is small. Due to the same reason, the subsamples of 9- and 10-year tenure are jointly used. The dash lines correspond to 95% confidence intervals. Standard errors are clustered within an establishment. 23 Table 1: Predictions and estimation results Hypotheses Data Information transmission Same Different Screening Peer effects Nepotism Outcome variables distribution distribution (1) (2) (3) (4) (5) (6) a. Job search spell − − − − b. Observable abilities − − c. Unobservable abilities + − 0 d. Tenure + + − e. Entry wage + − + + − f. Wage growth +/− 0 g. Productivity − + + − (−) Note. Column “Data” indicates our estimation results. + and − mean that use of social networks is, respectively, positively and negatively correlated. 0 means flat or no correlation. (−) is negative but statistically insignificant. A blank cell means no or ambiguous prediction. Table 2: Summary statistics: Establishments N % Industry Manufacturing 191 56 Commerce 74 22 Finance 74 22 Size Small (no. employees is 20 and less) 164 48 Medium (no. employees is 21-70) 99 29 Large (no. employees is 71 and above) 76 22 Single establishment firm 207 61 Social networks are a common mode of job advertisement 233 69 Note. N = 339. “Personal networks are a common mode of job advertisement ” indicates that an establishment answered that personal networks are a primary or secondary mode of advertising job vacancies. 24 Table 3: How jobs were found N % Social networks, including reference from somebody 2148 63 Family and relatives 773 23 Friends 931 27 Political affiliation 19 1 School alumni 30 1 Same village or town 395 12 Media advertisement and posting 1107 33 Through school 3 0 Public employment services 17 1 Private employment services 79 2 Job fairs 13 0 Internet posting 29 1 Observations 3396 Note. This table is based on the question in the survey, ”How did you find this job?” If a response is “social networks,” the survey further asked the type of networks used. 25 Table 4: Summary statistics: Employees All Social networks are used? Yes No Age at hire 24.21 23.56 25.33 (4.89) (5.05) (4.40) Schooing in years 9.01 7.19 12.15 (4.64) (3.84) (4.19) Tenure in years 4.32 4.23 4.47 (3.35) (3.36) (3.34) Earnings at hire (monthly, in Taka) 5745.66 4758.19 7445.25 (4213.63) (2865.85) (5443.01) Earnings (monthly, in Taka) 9146.08 7448.94 12067.13 (7108.49) (4418.45) (9510.92) Occupation at hire Manager 0.06 0.05 0.06 (0.23) (0.22) (0.23) Professional 0.20 0.16 0.28 (0.40) (0.37) (0.45) Semi-skilled worker 0.61 0.62 0.58 (0.49) (0.49) (0.49) Elementary job 0.14 0.16 0.08 (0.34) (0.37) (0.28) Search spell till finding the current job (in weeks) 8.60 5.56 13.85 (12.94) (5.91) (18.76) No. of vacancies applied for 3.60 2.28 5.87 (5.36) (2.79) (7.53) Promoted ever in the current firm (1=yes) 0.04 0.02 0.06 (0.18) (0.15) (0.23) Observations 3396 2148 1248 Note. Shown are means and standard deviations for permanent male employees aged 40 and below. Standard deviations are in parentheses. The second column is the statistics of those who found jobs through social networks, and the third is of those who did not. Earnings include bonuses, overtime, and other compensations. 26 Table 5: Job search spell and the no. of vacancies applied for: Comparing the employees who used social networks and those who did not Log(search spell) Log(no. applications) (1) (2) (3) (4) (5) (6) (7) (8) ∗∗∗ ∗∗ ∗∗ ∗∗ ∗∗∗ ∗∗ ∗∗ Social networks -0.437 -0.243 -0.202 -0.223 -0.428 -0.225 -0.203 -0.163 (0.086) (0.104) (0.096) (0.094) (0.099) (0.097) (0.092) (0.111) Schooing in years 0.069∗∗∗ 0.045∗∗∗ 0.022 0.043∗∗∗ 0.023∗∗∗ 0.023∗ (0.011) (0.010) (0.015) (0.012) (0.008) (0.012) Age at hire -0.013 -0.008 -0.010 0.013∗∗ 0.015∗∗∗ 0.013∗∗ (0.008) (0.009) (0.009) (0.006) (0.005) (0.006) Father’s pri.school 0.016 0.103 0.138∗ 0.034 0.091 0.076 (0.078) (0.079) (0.076) (0.068) (0.067) (0.076) Mother’s pri.school 0.045 0.126∗ 0.061 0.021 0.125∗∗ 0.137∗∗ (0.072) (0.066) (0.056) (0.061) (0.055) (0.067) Math z-score 0.004 0.011 -0.012 0.046 0.029 0.021 (0.034) (0.033) (0.044) (0.042) (0.035) (0.041) Language z-score -0.101∗∗ -0.100∗∗ -0.060 0.015 -0.014 -0.031 (0.042) (0.050) (0.054) (0.057) (0.058) (0.061) Big 5 index z-score -0.005 0.134 0.079 -0.072 -0.113∗∗ -0.133∗∗∗ (0.083) (0.100) (0.118) (0.051) (0.045) (0.047) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.150 0.213 0.159 0.116 0.100 0.166 0.119 0.078 N 3396 3390 3390 3390 3396 3390 3390 3390 Note. The variable, Social networks, is the dummy indicating that social networks were used to find jobs. The other controls included are entry year fixed effects and geographical division fixed effects. Occupations are grouped into manager, professional, semi-skilled, and elementary. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 27 Table 6: Observable abilities Schooling (years) Age at hire (1) (2) (3) (4) (5) (6) Social networks -3.401∗∗∗ -2.788∗∗∗ -1.394∗∗∗ -1.511∗∗∗ -1.673∗∗ -0.610 (0.315) (0.350) (0.270) (0.575) (0.680) (0.913) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.215 0.135 0.072 0.066 0.062 0.047 N 3396 3396 3396 3396 3396 3396 Pri. school: father Pri. school: mother (7) (8) (9) (10) (11) (12) Social networks -0.092∗∗ -0.059∗ -0.071∗∗ -0.160∗∗∗ -0.087∗∗ -0.089∗∗ (0.043) (0.034) (0.035) (0.041) (0.034) (0.039) Schooing in years 0.031∗∗∗ 0.026∗∗∗ 0.022∗∗∗ 0.032∗∗∗ 0.030∗∗∗ 0.018∗∗ (0.005) (0.004) (0.006) (0.006) (0.005) (0.007) Age at hire -0.002 -0.004 -0.005 0.006∗ 0.006∗∗ 0.007 (0.003) (0.003) (0.004) (0.003) (0.003) (0.004) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.160 0.078 0.044 0.190 0.112 0.050 N 3396 3396 3396 3396 3396 3396 Note. The dependent variables in columns (7)–(9) are the dummy indicating that the father completed primary education while those in columns (10)–(12) are the dummy for mother’s primary education. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 28 Table 7: Unobservable abilities Math (z-score) Language (z-score) (1) (2) (3) (4) (5) (6) (7) (8) Social networks -0.328∗∗∗ 0.057 0.042 0.049 -0.416∗∗∗ 0.107 0.120 0.066 (0.085) (0.077) (0.063) (0.073) (0.088) (0.087) (0.076) (0.089) Schooing in years 0.107∗∗∗ 0.101∗∗∗ 0.109∗∗∗ 0.129∗∗∗ 0.117∗∗∗ 0.148∗∗∗ (0.010) (0.012) (0.020) (0.011) (0.013) (0.016) Age at hire -0.006 -0.002 -0.003 -0.006 -0.003 0.000 (0.007) (0.007) (0.009) (0.006) (0.006) (0.006) Father’s pri.school 0.071 0.056 0.069 0.269∗∗∗ 0.164∗∗∗ 0.114∗ (0.083) (0.056) (0.058) (0.071) (0.058) (0.061) Mother’s pri.school 0.061 0.042 0.015 0.149 0.120 0.168 (0.090) (0.115) (0.123) (0.093) (0.100) (0.121) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.141 0.299 0.214 0.149 0.123 0.428 0.306 0.283 N 3396 3396 3396 3396 3396 3396 3396 3396 Big 5 (z-score) (9) (10) (11) (12) ∗ ∗∗ Social networks 0.069 0.132 0.034 0.002 (0.039) (0.052) (0.038) (0.040) Schooing in years 0.012 0.008 0.004 (0.009) (0.006) (0.006) Age at hire 0.004 -0.002 -0.003 (0.005) (0.005) (0.005) Father’s pri.school 0.075 -0.060∗ -0.051 (0.058) (0.032) (0.035) Mother’s pri.school 0.008 0.015 0.002 (0.066) (0.038) (0.035) Firm fixed effect X Firm-occ fixed effect X R-squared 0.109 0.124 0.053 0.058 N 3390 3390 3390 3390 Note. All the dependent variables are normalized to have the zero means and the one standard deviations. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 29 Table 8: Job tenure Log (tenure in years) (1) (2) (3) (4) Social networks -0.152∗ -0.166∗ -0.122 -0.197∗ (0.090) (0.099) (0.100) (0.109) Schooing in years -0.008 -0.004 -0.009 (0.008) (0.009) (0.010) Age at hire -0.010 -0.012∗ -0.005 (0.006) (0.007) (0.006) Father’s pri.school -0.019 -0.042 -0.072 (0.060) (0.063) (0.062) Mother’s pri.school 0.040 0.018 0.049 (0.052) (0.059) (0.072) Math z-score -0.020 0.034 0.035 (0.043) (0.056) (0.058) Language z-score 0.071 0.060 0.040 (0.053) (0.052) (0.057) Big 5 index z-score 0.066 0.085 0.158∗ (0.055) (0.082) (0.085) Firm fixed effect X Firm-occ fixed effect X R-squared 0.013 0.024 0.018 0.026 N 3396 3390 3390 3390 Note. The dependent variable log(tenure) is constructed by treating 0 year of tenure as 0.5 year, 1 year as 1.5 year, and so forth. The other controls included are establishment-entry occupation fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 30 Table 9: Entry salary Log (entry salary) (1) (2) (3) (4) Social networks -0.298∗∗∗ -0.070∗ -0.065∗ -0.089∗∗ (0.048) (0.042) (0.038) (0.043) Schooing in years 0.057∗∗∗ 0.061∗∗∗ 0.034∗∗∗ (0.006) (0.009) (0.008) Age at hire 0.022∗∗∗ 0.020∗∗∗ 0.013∗∗∗ (0.004) (0.003) (0.003) Father’s pri.school 0.052 0.058∗∗ 0.020 (0.033) (0.029) (0.027) Mother’s pri.school 0.043 0.056∗ 0.016 (0.035) (0.033) (0.037) Math z-score 0.069∗ 0.087∗ 0.089∗ (0.039) (0.046) (0.049) Language z-score -0.101∗∗∗ -0.098∗∗ -0.081∗ (0.036) (0.046) (0.049) Big 5 index z-score 0.021 -0.012 -0.082∗∗ (0.045) (0.032) (0.036) Firm fixed effect X Firm-occ fixed effect X R-squared 0.174 0.391 0.387 0.220 N 3396 3390 3390 3390 Note. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 31 Table 10: Current salary Log (current salary) (1) (2) (3) (4) Social networks -0.358∗∗∗ -0.129∗∗∗ -0.086∗∗∗ -0.118∗∗∗ (0.050) (0.036) (0.030) (0.032) Social networks*tenure 0.011 0.011 0.004 0.005 (0.007) (0.007) (0.005) (0.005) Tenure in years 0.032∗∗∗ 0.070∗∗∗ 0.077∗∗∗ 0.067∗∗∗ (0.006) (0.017) (0.016) (0.023) Abilities X X X Firm fixed effect X Firm-occ fixed effect X R-squared 0.222 0.493 0.520 0.377 N 3396 3390 3390 3390 Note. Abilities indicates that the observable and unobservable abilities (years of schooling, age at hire, and father’s and mother’s completion of primary education, math and language z-scores and Big 5 index) and their interactions with tenure are included or not. The other controls included are geographical division fixed effects. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. Table 11: Productivity: Promotion Dummy for having been promoted (1) (2) (3) (4) (5) (6) (7) (8) Social networks -0.043∗∗ -0.023 -0.017 -0.025 -0.001 -0.026 -0.003 -0.015 (0.019) (0.024) (0.024) (0.022) (0.032) (0.033) (0.031) (0.031) Social networks*tenure -0.009 0.002 -0.001 0.000 (0.007) (0.007) (0.007) (0.007) Tenure in years 0.016∗∗∗ 0.016∗∗∗ 0.015∗∗∗ 0.014∗∗∗ 0.021∗∗∗ 0.001 0.003 -0.001 (0.003) (0.003) (0.003) (0.003) (0.006) (0.018) (0.018) (0.021) Abilities X X X X X X Abilities*tenure X X X Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.077 0.088 0.091 0.110 0.081 0.122 0.137 0.167 N 3396 3390 3390 3390 3396 3390 3390 3390 Note. Abilities indicates that the observable and unobservable abilities (years of schooling, age at hire, and father’s and mother’s completion of primary education, math and language z-scores and Big 5 index) and their interactions with tenure are included or not. The other controls included are geographical division fixed effects. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 32 Table 12: Recruitment: Comparing the employers who use social networks and those who do not Employers use social networks as a common mode of job Diff. in advertisement? mean Yes No % employees who used social networks 82.859 36.962 45.897 (3.978)∗∗∗ N 233 106 Weeks to fill vacancies 2.611 4.813 -2.201 (0.492)∗∗∗ N 232 105 Difficulty in filling vacancies (normalized) -0.225 0.297 -0.522 (0.136)∗∗∗ N 227 98 No. vacancies posted in the last 12 months 10.443 9.718 0.725 (4.046) N 233 106 % vacancies filled in the last 12 months 96.523 97.340 -0.817 (1.827) N 130 51 Note. Means are shown above. The figures in parentheses are robust standard errors for the differences in means. The variable, Difficulty in filling vacancies, is based on responses to the question, “How difficult is it to fill vacancy?” asked for each occupation categories. The original answers are on the scale of 1–10 with 10 being the most difficult. We normalize them to have the zero means and the one standard deviations. 33 Table 13: Main reason for having chosen the job Good salary Good location Good work conditions (1) (2) (3) (4) (5) (6) Social networks 2.194 -0.075 1.967 0.298 -3.493 1.681 (2.210) (2.848) (3.953) (6.851) (3.129) (6.271) Abilities X X X Firm-occ fixed effect X X X R-squared 0.058 0.073 0.015 0.036 0.049 0.059 N 3394 3388 3394 3388 3394 3388 Control mean 19.359 19.359 11.718 11.718 32.621 32.621 Good prospect for Relevant to my Good prestige of career progression education firm (7) (8) (9) (10) (11) (12) Social networks -13.431∗∗∗ -8.901∗ -0.465 -0.251 -2.470∗∗ -2.770 (3.193) (4.791) (0.368) (0.253) (1.174) (1.761) Abilities X X X Firm-occ fixed effect X X X R-squared 0.076 0.066 0.023 0.022 0.047 0.084 N 3394 3388 3394 3388 3394 3388 Control mean 21.619 21.619 0.516 0.516 4.914 4.914 Recommended by No other offers others (13) (14) (15) (16) Social networks 6.645∗∗∗ 7.690∗∗∗ 9.053∗∗∗ 2.329 (1.475) (2.716) (2.841) (2.604) Abilities X X Firm-occ fixed effect X X R-squared 0.046 0.049 0.049 0.046 N 3394 3388 3394 3388 Control mean 0.721 0.721 8.533 8.533 Note. This table presents linear probability models analyzing main reasons why employees chose the jobs. The dependent variable is the binary dummy for the main reason being the one indicated in a corresponding column header. The survey asked employees which of listed reasons was the main reason. The row Control mean shows the sample means of the dependent variables among the employees who did not use social networks. The row Abilities indicates whether the observable and unobservable characteristics (years of schooling, age at hire, and father’s and mother’s years of schooling, math and language z-scores, and Big 5 index) are included or not. All specifications include entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Coefficients, standard errors, and control means are multiplied by 100 and should be interpreted as percentage points. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 34 Table 14: Who used social networks? Social network dummy (1) (2) (3) (4) Schooing in years -0.045∗∗∗ -0.050∗∗∗ -0.029∗∗∗ -0.025∗∗∗ (0.003) (0.003) (0.003) (0.004) Age at hire -0.004∗∗ -0.006∗∗∗ -0.005∗∗∗ -0.005∗∗∗ (0.002) (0.002) (0.002) (0.002) Father’s pri.school -0.014 -0.011 -0.039∗ -0.034 (0.025) (0.025) (0.021) (0.023) Mother’s pri.school -0.081∗∗∗ -0.081∗∗∗ -0.041∗∗ -0.037 (0.022) (0.022) (0.021) (0.023) Math z-score -0.011 -0.012 0.003 0.001 (0.014) (0.014) (0.014) (0.016) Language z-score -0.007 -0.004 0.002 -0.019 (0.015) (0.015) (0.014) (0.016) Big 5 index z-score 0.066∗∗∗ 0.061∗∗ 0.030 0.027 (0.024) (0.024) (0.020) (0.021) Entry occupation fixed effect X Firm fixed effect X Firm-occ fixed effect X R-squared 0.288 0.299 0.098 0.056 N 3390 3390 3390 3390 Note. All specifications control for geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 35 A Appendix: Robustness check on attrition bias 36 Log of search spell Log of no. of applications 2 1 Coefficient of the social network dummy Coefficient of the social network dummy 1 0 0 -1 -1 -2 -2 -3 -3 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9 Tenure in years Tenure in years (a) Job search spell (b) No. of vacancies applied for Years of schooling Father completed pri.school .5 6 Coefficient of the social network dummy Coefficient of the social network dummy 4 0 2 -.5 0 -1 -2 -1.5 -4 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9 Tenure in years Tenure in years (c) Years of schooling (d) Father’s completion of pri. education Mother completed pri.school Log of entry wage 1.5 1 Coefficient of the social network dummy Coefficient of the social network dummy .5 .5 1 -.5 0 0 -1 -.5 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9 Tenure in years Tenure in years (e) Mother’s completion of pri. education (f) Log of entry salary Dummy for being ever promoted 1 Coefficient of the social network dummy -.5 0 -1 .5 1 2 3 4 5 6 7 8 9 Tenure in years (g) Promotion Figure A1: Robustness check—Comparison at each year of job tenure Note. Plotted are estimates of θ in the equation of yijk = θN etworkijk + β Xijk + γjk + ijk . In figures (a), (b), (f), and (g), Xijk are years of schooling, age at hire, math and language z-scores, Big 5 index, and the dummies for father’s and mother’s completion of primary education. In figures (d) and (e), Xijk are years of schooling and age at hire. In figure (c), Xijk are not included for. The regressions are run separately for a subsample of each year of tenure, except for 1-year and 9-year tenure. The subsamples of 0- and 1-year tenure are jointly used because their size is small. Due to the same reason, the subsamples of 9- and 10-year tenure are jointly used. The dash lines are 95% confidence intervals. Standard errors are clustered within an establishment. 37 Table A1: Robustness check using the subsample with tenure being 3 years and less—Job tenure Log (tenure in years) (1) (2) (3) (4) Social networks -0.083 -0.039 -0.040 -0.091∗∗ (0.065) (0.054) (0.041) (0.040) Schooing in years -0.009 -0.010 -0.003 (0.007) (0.008) (0.009) Age at hire 0.001 -0.004 -0.001 (0.006) (0.006) (0.006) Father’s pri.school 0.080 -0.036 -0.065 (0.082) (0.072) (0.075) Mother’s pri.school 0.087 0.128∗∗ 0.172∗∗∗ (0.054) (0.055) (0.056) Math z-score -0.068∗ -0.015 0.022 (0.038) (0.027) (0.020) Language z-score 0.113∗∗ 0.103∗∗ 0.048 (0.054) (0.048) (0.047) Big 5 index z-score 0.014 -0.043 -0.085 (0.051) (0.068) (0.066) Firm fixed effect X Firm-occ fixed effect X R-squared 0.015 0.063 0.041 0.048 N 1745 1744 1744 1744 Note. The dependent variable log(tenure) is constructed by treating 0 year of tenure as 0.5 year, 1 year as 1.5 year, and so forth. The other controls included are establishment-entry occupation fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 38 Table A2: Robustness check using the subsample with tenure being 3 years and less—Job search spell and the no. of vacancies applied for Log(search spell) Log(no. applications) (1) (2) (3) (4) (5) (6) (7) (8) ∗∗∗ ∗∗∗ ∗ ∗ ∗∗∗ ∗∗∗ ∗∗∗ Social networks -0.464 -0.322 -0.236 -0.318 -0.434 -0.257 -0.232 -0.241∗∗ (0.084) (0.106) (0.122) (0.165) (0.101) (0.088) (0.088) (0.099) Schooing in years 0.059∗∗∗ 0.047∗∗∗ 0.041∗∗ 0.038∗∗∗ 0.020 0.035∗∗ (0.012) (0.010) (0.019) (0.014) (0.015) (0.018) Age at hire -0.012 -0.010 -0.010 0.018∗∗∗ 0.025∗∗∗ 0.029∗∗∗ (0.010) (0.012) (0.011) (0.006) (0.005) (0.008) Father’s pri.school 0.091 0.188∗∗ 0.182∗∗ 0.055 0.129 0.100 (0.080) (0.084) (0.085) (0.108) (0.125) (0.153) Mother’s pri.school -0.007 0.049 0.014 -0.004 0.053 0.107 (0.062) (0.066) (0.069) (0.062) (0.062) (0.075) Math z-score -0.001 -0.043 -0.059 0.052 0.025 0.051 (0.051) (0.063) (0.093) (0.046) (0.048) (0.061) Language z-score -0.112∗∗ -0.105 -0.124 -0.021 -0.046 -0.139 (0.045) (0.068) (0.095) (0.065) (0.078) (0.098) Big 5 index z-score 0.039 0.182∗ 0.095 -0.064 -0.140∗∗ -0.203∗∗∗ (0.090) (0.110) (0.132) (0.058) (0.061) (0.070) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.129 0.176 0.103 0.095 0.093 0.153 0.115 0.108 N 1745 1744 1744 1744 1745 1744 1744 1744 Note. The variable, Social networks, is the dummy indicating that social networks were used to find jobs. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 39 Table A3: Robustness check using the subsample with tenure being 3 years and less— Observable abilities Schooling (years) Age at hire (1) (2) (3) (4) (5) (6) Social networks -3.444∗∗∗ -2.846∗∗∗ -1.521∗∗∗ -1.553∗∗ -1.695∗ -0.339 (0.349) (0.416) (0.406) (0.738) (0.928) (1.491) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.237 0.151 0.079 0.034 0.029 0.004 N 1745 1745 1745 1745 1745 1745 Pri. school: father Pri. school: mother (7) (8) (9) (10) (11) (12) Social networks -0.047 -0.030 -0.041 -0.184∗∗∗ -0.124∗∗ -0.096∗∗ (0.039) (0.037) (0.041) (0.047) (0.050) (0.044) Schooing in years 0.035∗∗∗ 0.027∗∗∗ 0.025∗∗∗ 0.032∗∗∗ 0.023∗∗ 0.017 (0.005) (0.006) (0.007) (0.007) (0.009) (0.011) Age at hire -0.006 -0.006 -0.005 -0.001 0.003 -0.001 (0.005) (0.006) (0.008) (0.004) (0.005) (0.006) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.163 0.058 0.033 0.184 0.089 0.053 N 1745 1745 1745 1745 1745 1745 Note. The dependent variables in columns (7)–(9) are the dummy indicating that the father completed primary education while those in columns (10)–(12) are the dummy for mother’s primary education. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 40 Table A4: Robustness check using the subsample with tenure being 3 years and less— Unobservable abilities Math (z-score) Language (z-score) (1) (2) (3) (4) (5) (6) (7) (8) Social networks -0.293∗∗∗ 0.090 0.171∗ 0.235∗ -0.463∗∗∗ 0.060 0.061 -0.001 (0.103) (0.105) (0.092) (0.122) (0.079) (0.076) (0.066) (0.071) Schooing in years 0.115∗∗∗ 0.113∗∗∗ 0.133∗∗∗ 0.134∗∗∗ 0.116∗∗∗ 0.146∗∗∗ (0.014) (0.015) (0.027) (0.014) (0.017) (0.021) Age at hire -0.006 0.002 -0.003 -0.011∗ -0.005 -0.002 (0.011) (0.012) (0.013) (0.007) (0.009) (0.007) Father’s pri.school -0.015 0.140 0.171∗∗ 0.210∗∗ 0.183∗∗ 0.136 (0.111) (0.094) (0.082) (0.087) (0.088) (0.089) Mother’s pri.school -0.005 -0.089 -0.180 0.158 0.047 0.072 (0.138) (0.170) (0.174) (0.123) (0.117) (0.149) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.125 0.270 0.213 0.169 0.149 0.440 0.293 0.276 N 1745 1745 1745 1745 1745 1745 1745 1745 Big 5 (z-score) (9) (10) (11) (12) ∗ Social networks 0.063 0.098 0.043 -0.032 (0.038) (0.063) (0.057) (0.065) Schooing in years 0.007 0.001 -0.008 (0.009) (0.008) (0.007) Age at hire 0.002 -0.005 -0.006 (0.007) (0.007) (0.006) Father’s pri.school 0.053 -0.056 -0.061 (0.084) (0.062) (0.063) Mother’s pri.school -0.003 -0.003 -0.001 (0.084) (0.056) (0.047) Firm fixed effect X Firm-occ fixed effect X R-squared 0.122 0.128 0.028 0.029 N 1744 1744 1744 1744 Note. All the dependent variables are normalized to have the zero means and the one standard deviations. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 41 Table A5: Robustness check using the subsample with tenure being 3 years and less— Entry salary Log (entry salary) (1) (2) (3) (4) Social networks -0.355∗∗∗ -0.103∗∗∗ -0.083∗ -0.079∗∗ (0.046) (0.038) (0.045) (0.036) Schooing in years 0.065∗∗∗ 0.076∗∗∗ 0.053∗∗∗ (0.006) (0.008) (0.009) Age at hire 0.029∗∗∗ 0.025∗∗∗ 0.015∗∗∗ (0.004) (0.004) (0.004) Father’s pri.school 0.034 0.022 0.027 (0.042) (0.039) (0.036) Mother’s pri.school 0.029 0.092∗∗ 0.011 (0.037) (0.041) (0.042) Math z-score 0.068∗ 0.071∗∗ 0.049∗ (0.035) (0.036) (0.028) Language z-score -0.105∗∗ -0.090∗ -0.036 (0.042) (0.046) (0.042) Big 5 index z-score 0.028 0.006 -0.020 (0.040) (0.039) (0.036) Firm fixed effect X Firm-occ fixed effect X R-squared 0.174 0.489 0.521 0.289 N 1745 1744 1744 1744 Note. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 42 Table A6: Robustness check using the subsample with tenure being 3 years and less— Current salary Log (current salary) (1) (2) (3) (4) Social networks -0.477∗∗∗ -0.312∗∗∗ -0.328∗∗∗ -0.257∗ (0.169) (0.119) (0.113) (0.143) Social networks*tenure 0.055 0.085∗ 0.100∗∗ 0.063 (0.062) (0.045) (0.044) (0.048) Tenure in years 0.026 0.047 -0.001 0.004 (0.068) (0.130) (0.115) (0.112) Abilities X X X Firm fixed effect X Firm-occ fixed effect X R-squared 0.179 0.511 0.555 0.312 N 1745 1744 1744 1744 Note. Abilities indicates that the observable and unobservable abilities (years of schooling, age at hire, and father’s and mother’s completion of primary education, math and language z-scores and Big 5 index) and their interactions with tenure are included or not. The other controls included are geographical division fixed effects. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. Table A7: Robustness check using the subsample with tenure being 3 years and less— Promotion Dummy for having been promoted (1) (2) (3) (4) (5) (6) (7) (8) Social networks -0.028 -0.021 -0.016 -0.015 -0.053 -0.075 -0.061 -0.069 (0.022) (0.025) (0.021) (0.023) (0.052) (0.059) (0.040) (0.054) Social networks*tenure 0.010 0.022 0.019∗ 0.023 (0.014) (0.016) (0.011) (0.016) Tenure in years 0.002 0.001 0.015∗∗ 0.016∗ -0.005 -0.022 -0.018 -0.035 (0.005) (0.006) (0.007) (0.008) (0.014) (0.016) (0.014) (0.024) Abilities X X X X X X Abilities*tenure X X X Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.016 0.034 0.045 0.072 0.017 0.050 0.054 0.087 N 1745 1744 1744 1744 1745 1744 1744 1744 Note. Abilities indicates that the observable and unobservable abilities (years of schooling, age at hire, and father’s and mother’s completion of primary education, math and language z-scores and Big 5 index) and their interactions with tenure are included or not. The other controls included are geographical division fixed effects. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 43 Table A8: Robustness check using the subsample with tenure being 3 years and less— Main reason for choosing the job Good salary Good location Good work conditions (1) (2) (3) (4) (5) (6) Social networks 3.901 -0.870 -3.737 -5.390 -5.892 1.052 (2.742) (3.543) (6.337) (8.460) (5.814) (8.817) Abilities X X X Firm-occ fixed effect X X X R-squared 0.063 0.034 0.016 0.077 0.016 0.073 N 1744 1743 1744 1743 1744 1743 Control mean 21.558 21.558 14.974 14.974 32.320 32.320 Good prospect for Relevant to my Good prestige of career progression education firm (7) (8) (9) (10) (11) (12) Social networks -9.517∗∗∗ -5.888 -0.027 -0.062 -0.343 -0.186 (3.329) (5.701) (0.083) (0.053) (1.167) (1.574) Abilities X X X Firm-occ fixed effect X X X R-squared 0.049 0.055 0.003 0.001 0.084 0.028 N 1744 1743 1744 1743 1744 1743 Control mean 16.472 16.472 0.102 0.102 4.819 4.819 Recommended by No other offers others (13) (14) (15) (16) Social networks 7.388∗∗∗ 8.589∗∗ 8.227∗∗∗ 2.756 (1.990) (3.850) (2.681) (3.078) Abilities X X Firm-occ fixed effect X X R-squared 0.039 0.059 0.038 0.055 N 1744 1743 1744 1743 Control mean 1.178 1.178 8.578 8.578 Note. This table presents linear probability models analyzing main reasons why employees chose the jobs. The dependent variable is the binary dummy for the main reason being the one indicated in a corresponding column header. The survey asked employees which of listed reasons was the main reason. The row Control mean shows the sample means of the dependent variables among the employees who did not use social networks. The row Abilities indicates whether the observable and unobservable characteristics (years of schooling, age at hire, and father’s and mother’s years of schooling, math and language z-scores, and Big 5 index) are included or not. All specifications include geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Coefficients, standard errors, and control means are multiplied by 100 and should be interpreted as percentage points. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 44 B Appendix: Robustness check on cutoff age to re- strict the sample Table B1: Robustness check using the sample of the aged 60 and below—Job search Log(search spell) Log(no. applications) (1) (2) (3) (4) (5) (6) (7) (8) ∗∗∗ ∗∗∗ ∗∗ ∗∗ ∗∗∗ ∗∗ ∗∗ Social networks -0.465 -0.274 -0.229 -0.225 -0.452 -0.242 -0.226 -0.187∗ (0.087) (0.105) (0.099) (0.093) (0.098) (0.093) (0.089) (0.105) Schooing in years 0.065∗∗∗ 0.038∗∗∗ 0.018 0.048∗∗∗ 0.026∗∗∗ 0.027∗∗ (0.011) (0.010) (0.013) (0.012) (0.008) (0.011) Age at hire -0.009 -0.003 -0.004 0.007∗∗ 0.010∗∗∗ 0.008∗∗ (0.006) (0.006) (0.006) (0.004) (0.003) (0.004) Father’s pri.school 0.034 0.130∗ 0.150∗∗ 0.039 0.089 0.062 (0.079) (0.078) (0.073) (0.064) (0.060) (0.067) Mother’s pri.school 0.026 0.103∗ 0.029 0.023 0.129∗∗ 0.141∗∗ (0.066) (0.060) (0.049) (0.056) (0.054) (0.062) Math z-score 0.005 0.012 -0.014 0.036 0.021 0.014 (0.034) (0.033) (0.042) (0.040) (0.032) (0.038) Language z-score -0.101∗∗ -0.086∗ -0.058 0.007 -0.006 -0.029 (0.041) (0.048) (0.050) (0.054) (0.053) (0.059) Big 5 index z-score 0.001 0.144 0.081 -0.064 -0.113∗∗ -0.135∗∗∗ (0.087) (0.107) (0.121) (0.049) (0.046) (0.050) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.166 0.223 0.147 0.108 0.115 0.182 0.131 0.094 N 3763 3757 3757 3757 3763 3757 3757 3757 Note. The variable, Social networks, is the dummy indicating that social networks were used to find jobs. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 45 Table B2: Robustness check using the sample of the aged 60 and below—Observable abilities Schooling (years) Age at hire (1) (2) (3) (4) (5) (6) Social networks -3.515∗∗∗ -2.851∗∗∗ -1.382∗∗∗ -1.146∗ -1.304∗ -0.070 (0.311) (0.338) (0.263) (0.667) (0.782) (0.891) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.222 0.146 0.095 0.059 0.067 0.071 N 3763 3763 3763 3763 3763 3763 Pri. school: father Pri. school: mother (7) (8) (9) (10) (11) (12) Social networks -0.092∗∗ -0.054∗ -0.062∗ -0.157∗∗∗ -0.085∗∗∗ -0.081∗∗ (0.042) (0.033) (0.034) (0.040) (0.033) (0.040) Schooing in years 0.031∗∗∗ 0.028∗∗∗ 0.024∗∗∗ 0.033∗∗∗ 0.034∗∗∗ 0.024∗∗∗ (0.004) (0.004) (0.006) (0.005) (0.005) (0.007) Age at hire -0.002 -0.004∗ -0.004∗ 0.002 0.001 0.000 (0.003) (0.002) (0.002) (0.002) (0.002) (0.003) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.167 0.099 0.065 0.189 0.118 0.062 N 3763 3763 3763 3763 3763 3763 Note. The dependent variables in columns (7)–(9) are the dummy indicating that the father completed primary education while those in columns (10)–(12) are the dummy for mother’s primary education. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 46 Table B3: Robustness check using the sample of the aged 60 and below—Unobservable abilities Math (z-score) Language (z-score) (1) (2) (3) (4) (5) (6) (7) (8) Social networks -0.328∗∗∗ 0.069 0.057 0.060 -0.433∗∗∗ 0.102 0.109 0.044 (0.084) (0.070) (0.062) (0.070) (0.088) (0.078) (0.067) (0.079) Schooing in years 0.103∗∗∗ 0.101∗∗∗ 0.111∗∗∗ 0.127∗∗∗ 0.116∗∗∗ 0.143∗∗∗ (0.010) (0.010) (0.016) (0.010) (0.012) (0.014) Age at hire 0.004 0.002 0.002 -0.002 -0.002 0.001 (0.005) (0.005) (0.006) (0.003) (0.004) (0.004) Father’s pri.school 0.096 0.075 0.081 0.266∗∗∗ 0.161∗∗∗ 0.103∗ (0.078) (0.055) (0.058) (0.070) (0.059) (0.059) Mother’s pri.school 0.043 0.022 0.005 0.133 0.109 0.145 (0.085) (0.110) (0.122) (0.084) (0.090) (0.110) Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.138 0.298 0.226 0.166 0.127 0.431 0.318 0.288 N 3763 3763 3763 3763 3763 3763 3763 3763 Big 5 (z-score) (9) (10) (11) (12) ∗ ∗∗ Social networks 0.063 0.123 0.030 -0.005 (0.038) (0.051) (0.042) (0.043) Schooing in years 0.011 0.007 0.002 (0.008) (0.006) (0.004) Age at hire 0.003 -0.001 -0.000 (0.004) (0.004) (0.004) Father’s pri.school 0.077 -0.054∗ -0.044 (0.057) (0.030) (0.032) Mother’s pri.school 0.006 0.021 -0.002 (0.068) (0.037) (0.038) Firm fixed effect X Firm-occ fixed effect X R-squared 0.107 0.121 0.046 0.054 N 3757 3757 3757 3757 Note. All the dependent variables are normalized to have the zero means and the one standard deviations. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 47 Table B4: Robustness check using the sample of the aged 60 and below—Job tenure Log (tenure in years) (1) (2) (3) (4) Social networks -0.134∗ -0.120 -0.068 -0.135 (0.080) (0.093) (0.098) (0.115) Schooing in years -0.004 -0.004 -0.009 (0.008) (0.009) (0.010) Age at hire 0.001 0.003 0.007 (0.005) (0.006) (0.006) Father’s pri.school -0.061 -0.092 -0.115∗ (0.064) (0.068) (0.064) Mother’s pri.school 0.046 -0.007 0.036 (0.053) (0.056) (0.070) Math z-score -0.011 0.033 0.036 (0.045) (0.058) (0.062) Language z-score 0.076 0.065 0.061 (0.058) (0.059) (0.073) Big 5 index z-score 0.053 0.060 0.114 (0.060) (0.082) (0.085) Firm fixed effect X Firm-occ fixed effect X R-squared 0.014 0.022 0.011 0.020 N 3763 3757 3757 3757 Note. The dependent variable log(tenure) is constructed by treating 0 year of tenure as 0.5 year, 1 year as 1.5 year, and so forth. The other controls included are establishment-entry occupation fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 48 Table B5: Robustness check using the sample of the aged 60 and below—Entry salary Log (entry salary) (1) (2) (3) (4) Social networks -0.304∗∗∗ -0.071∗ -0.067∗∗ -0.081∗∗ (0.049) (0.038) (0.033) (0.036) Schooing in years 0.059∗∗∗ 0.062∗∗∗ 0.038∗∗∗ (0.005) (0.008) (0.006) Age at hire 0.018∗∗∗ 0.016∗∗∗ 0.011∗∗∗ (0.003) (0.003) (0.003) Father’s pri.school 0.062∗ 0.062∗∗ 0.020 (0.033) (0.029) (0.030) Mother’s pri.school 0.064∗ 0.069∗∗ 0.041 (0.033) (0.032) (0.038) Math z-score 0.065 0.086∗ 0.094∗ (0.040) (0.048) (0.051) Language z-score -0.105∗∗∗ -0.093∗∗ -0.086∗ (0.036) (0.046) (0.050) Big 5 index z-score 0.005 -0.030 -0.093∗∗ (0.045) (0.036) (0.042) Firm fixed effect X Firm-occ fixed effect X R-squared 0.207 0.410 0.417 0.299 N 3763 3757 3757 3757 Note. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 49 Table B6: Robustness check using the sample of the aged 60 and below—Current salary Log (current salary) (1) (2) (3) (4) Social networks -0.351∗∗∗ -0.112∗∗∗ -0.072∗∗ -0.083∗∗∗ (0.049) (0.034) (0.030) (0.025) Social networks*tenure 0.008 0.005 -0.001 -0.000 (0.006) (0.006) (0.005) (0.006) Tenure in years 0.030∗∗∗ 0.073∗∗∗ 0.082∗∗∗ 0.076∗∗∗ (0.003) (0.016) (0.014) (0.016) Abilities X X X Firm fixed effect X Firm-occ fixed effect X R-squared 0.223 0.493 0.523 0.407 N 3763 3757 3757 3757 Note. Abilities indicates that the observable and unobservable abilities (years of schooling, age at hire, and father’s and mother’s completion of primary education, math and language z-scores and Big 5 index) and their interactions with tenure are included or not. The other controls included are geographical division fixed effects. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. Table B7: Robustness check using the sample of the aged 60 and below—Promotion Dummy for having been promoted (1) (2) (3) (4) (5) (6) (7) (8) Social networks -0.028 -0.021 -0.016 -0.015 -0.053 -0.075 -0.061 -0.069 (0.022) (0.025) (0.021) (0.023) (0.052) (0.059) (0.040) (0.054) Social networks*tenure 0.010 0.022 0.019∗ 0.023 (0.014) (0.016) (0.011) (0.016) Tenure in years 0.002 0.001 0.015∗∗ 0.016∗ -0.005 -0.022 -0.018 -0.035 (0.005) (0.006) (0.007) (0.008) (0.014) (0.016) (0.014) (0.024) Abilities X X X X X X Abilities*tenure X X X Firm fixed effect X X Firm-occ fixed effect X X R-squared 0.016 0.034 0.045 0.072 0.017 0.050 0.054 0.087 N 1745 1744 1744 1744 1745 1744 1744 1744 Note. Abilities indicates that the observable and unobservable abilities (years of schooling, age at hire, and father’s and mother’s completion of primary education, math and language z-scores and Big 5 index) and their interactions with tenure are included or not. The other controls included are geographical division fixed effects. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 50 Table B8: Robustness check using the sample of the aged 60 and below—Main reason for having chosen the job Good salary Good location Good work conditions (1) (2) (3) (4) (5) (6) Social networks 1.393 -0.351 1.940 -0.501 -2.753 2.062 (2.190) (2.830) (3.708) (6.318) (2.818) (5.623) Abilities X X X Firm-occ fixed effect X X X R-squared 0.064 0.082 0.030 0.052 0.045 0.054 N 3761 3755 3761 3755 3761 3755 Control mean 20.157 20.157 12.380 12.380 31.604 31.604 Good prospect for Relevant to my Good prestige of career progression education firm (7) (8) (9) (10) (11) (12) Social networks -12.834∗∗∗ -7.790 -0.378 -0.228 -2.467∗∗ -2.702 (3.440) (5.175) (0.347) (0.225) (1.103) (1.776) Abilities X X X Firm-occ fixed effect X X X R-squared 0.072 0.070 0.036 0.044 0.041 0.076 N 3761 3755 3761 3755 3761 3755 Control mean 21.456 21.456 0.512 0.512 4.791 4.791 Recommended by No other offers others (13) (14) (15) (16) Social networks 6.510∗∗∗ 7.642∗∗∗ 8.588∗∗∗ 1.870 (1.413) (2.432) (2.647) (2.400) Abilities X X Firm-occ fixed effect X X R-squared 0.053 0.060 0.048 0.048 N 3761 3755 3761 3755 Control mean 0.760 0.760 8.340 8.340 Note. This table presents linear probability models analyzing main reasons why employees chose the jobs. The dependent variable is the binary dummy for the main reason being the one indicated in a corresponding column header. The survey asked employees which of listed reasons was the main reason. The row Control mean shows the sample means of the dependent variables among the employees who did not use social networks. The row Abilities indicates whether the observable and unobservable characteristics (years of schooling, age at hire, and father’s and mother’s years of schooling, math and language z-scores, and Big 5 index) are included or not. All specifications include geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Coefficients, standard errors, and control means are multiplied by 100 and should be interpreted as percentage points. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 51 C Appendix: Big 5 and social networks . Table C1: Comparing individual Big 5 personality between the employees who used social networks and who did not Conscientiousness Emotional stability Agreeableness (1) (2) (3) (4) (5) (6) Social networks 0.138 0.047 0.042 -0.009 -0.009 -0.072 (0.130) (0.097) (0.061) (0.048) (0.063) (0.078) Schooing in years 0.024∗∗ 0.015 0.006 (0.011) (0.015) (0.020) Age at hire 0.002 0.008 -0.014 (0.005) (0.007) (0.010) Father’s pri.school -0.083 -0.029 0.022 (0.098) (0.065) (0.060) Mother’s pri.school 0.009 0.024 -0.028 (0.054) (0.062) (0.083) Firm-occ fixed effect X X X R-squared 0.038 0.040 0.022 0.043 0.031 0.057 N 3386 3386 3346 3346 3340 3340 Extraversion Openness (7) (8) (9) (10) Social networks 0.086 -0.055 0.017 0.091 (0.063) (0.109) (0.077) (0.064) Schooing in years -0.013 -0.014 (0.020) (0.013) Age at hire -0.002 -0.006 (0.009) (0.010) Father’s pri.school -0.156∗ -0.004 (0.090) (0.051) Mother’s pri.school -0.091 0.097∗ (0.101) (0.050) Firm-occ fixed effect X X R-squared 0.042 0.060 0.028 0.039 N 3377 3377 3383 3383 Note. All the dependent variables are normalized to have the zero means and the one standard deviations. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 52 Table C2: Individual Big 5 personality: Job search spells, no. of applications, tenure, entry salary Log(search Log(no. Log(tenure in Log(entry spell) applications) years) salary) (1) (2) (3) (4) (5) (6) (7) (8) Social networks -0.242∗∗ -0.223∗∗∗ -0.232∗∗ -0.168 -0.161 -0.184∗ -0.069 -0.085∗∗ (0.098) (0.083) (0.101) (0.118) (0.098) (0.106) (0.042) (0.040) Schooing in years 0.066∗∗∗ 0.021 0.041∗∗∗ 0.022∗ -0.007 -0.007 0.056∗∗∗ 0.033∗∗∗ (0.011) (0.014) (0.012) (0.012) (0.008) (0.011) (0.006) (0.008) Age at hire -0.012∗ -0.009 0.014∗∗ 0.014∗∗ -0.010 -0.007 0.023∗∗∗ 0.013∗∗∗ (0.007) (0.008) (0.006) (0.006) (0.007) (0.007) (0.004) (0.003) Father’s pri.school 0.029 0.135∗ 0.040 0.077 -0.017 -0.064 0.057∗ 0.022 (0.082) (0.077) (0.069) (0.073) (0.060) (0.059) (0.033) (0.027) Mother’s pri.school 0.030 0.053 0.030 0.132∗ 0.052 0.055 0.046 0.021 (0.067) (0.058) (0.058) (0.071) (0.054) (0.069) (0.035) (0.040) Math z-score 0.017 -0.008 0.062 0.031 -0.022 0.030 0.075∗ 0.096∗ (0.034) (0.041) (0.041) (0.042) (0.043) (0.050) (0.042) (0.049) Language z-score -0.091∗∗ -0.049 0.014 -0.020 0.072 0.029 -0.096∗∗∗ -0.082∗∗ (0.037) (0.046) (0.051) (0.057) (0.052) (0.054) (0.035) (0.040) Big 5 personality traits Conscientiousness -0.074∗∗ -0.020 -0.024 -0.072∗∗ 0.020 0.058 -0.029 -0.044∗ (0.036) (0.035) (0.028) (0.029) (0.023) (0.038) (0.018) (0.026) Emotional stability -0.020 -0.037 -0.055∗ -0.057∗ 0.042∗ 0.082∗∗ 0.026∗∗ 0.034∗∗ (0.031) (0.031) (0.030) (0.033) (0.023) (0.032) (0.013) (0.015) Agreeableness 0.026 0.098∗ -0.028 -0.024 -0.029 -0.035 0.007 -0.021 (0.036) (0.055) (0.036) (0.036) (0.020) (0.032) (0.015) (0.013) Extraversion 0.021 0.006 0.060∗ -0.008 0.045 0.061∗ 0.031 0.009 (0.032) (0.038) (0.033) (0.022) (0.033) (0.037) (0.021) (0.018) Openness 0.063∗∗ 0.039 -0.027 0.023 0.000 -0.006 0.005 -0.057∗∗∗ (0.026) (0.036) (0.042) (0.051) (0.035) (0.046) (0.020) (0.017) Firm-occ fixed effect X X X X R-squared 0.225 0.137 0.176 0.085 0.032 0.040 0.398 0.252 N 3315 3315 3315 3315 3315 3315 3315 3315 Note. The dependent variable log(tenure) is constructed by treating 0 year of tenure as 0.5 year, 1 year as 1.5 year, and so forth. The other controls included are entry year fixed effects and geographical division fixed effects. Standard errors clustered within an establishment are in parentheses. Significance levels: ∗∗∗ = 1%, ∗∗ = 5%, ∗ = 10%. 53