How did you find your job? Effects of the job search channels on labour market outcomes in Germany

Afonina, Mariya; Zaharieva, Anna Working Paper How did you find your job? Effects of the job search channels on labour market outcomes in Germany Working Papers in Economics and Management, No. 1-2025 Provided in Cooperation with: Faculty of Business Administration and Economics, Bielefeld University Suggested Citation: Afonina, Mariya; Zaharieva, Anna (2025) : How did you find your job? Effects of the job search channels on labour market outcomes in Germany, Working Papers in Economics and Management, No. 1-2025, Bielefeld University, Faculty of Business Administration and Economics, Bielefeld, https://doi.org/10.4119/unibi/3001180 This Version is available at: https://hdl.handle.net/10419/315015 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/    Faculty of Business Administration and Economics www.wiwi.uni−bielefeld.de 33501 Bielefeld − Germany P.O. Box 10 01 31 Bielefeld University ISSN 2196−2723 Working Papers in Economics and Management ➔ No. 1-2025 March 2025 How did you find your job? Effects of the job search channels on labour market outcomes in Germany Mariya Afonina, Anna Zaharieva How did you find your job? Effects of the job search channels on labour market outcomes in Germany ∗ Mariya Afonina † , Anna Zaharieva ‡ February 26, 2025 Abstract We study the effect of finding a job through one’s social contact on starting wages. Using combined SOEP-INKAR data for Germany and propensity score analysis - both matching and weighting - we document that referral hiring is associated with a wage penalty of 10%. This penalty is stable over time. Separating by the type of the social contact, we find that referrals from former colleagues are associated with a 9% wage premium compared to a direct formal application. In contrast, referrals from friends are associated with a 7% wage penalty. Our results highlight persistent self-selection of workers on observable and unobservable characteristics. Using information from a short test of cognitive abilities (symbol digit test) we document that workers recommended by former colleagues perform best in the ability test, consistent with the predictions from a sorting model. The lowest performance is recorded for those relying on the help of their friends. The effects are primarily driven by the sub-sample of women. No significant differences across search channels are found for personality traits. JEL-code: C21, J31, J62, J64 Key words: job search, social networks, referrals, cognitive abilities ∗ We thank Herbert Dawid, Jan Goebel, Zainab Iftikhar, Moritz Kuhn, Simon K¨uhne, Marcus Pannenberg and Katrin Rickmeier for their valuable comments. The authors also would like to thank the Leibniz Association for providing financial support to the Leibniz Science Campus “SOEPRegioHub” at Bielefeld University. † Bielefeld Graduate School of Economics and Management, Bielefeld University, 33501 Bielefeld, Germany. Email: [email protected] ‡ Chair for Labour Economics, Bielefeld University, 33501 Bielefeld, Germany. Email: anna.zahariev[email protected] 1 Introduction There are multiple ways to find a job varying from a formal application in Internet, intermediation by the employment agency to relying on the help from one’s social network. The literature suggests that finding a job through a social contact is the most frequent channel of entering a job making up 30 −50% of all new matches1. Recognizing the importance of this search channel for the labour market, multiple studies tried to evaluate the relationship between the job search channel, which generated a job, and the starting wage. The empirical evidence is mixed with equally large groups of studies supporting the idea of wage premia and wage penalties from referral hiring compared to direct formal applications2. This raises two questions, first, whether the selection of workers to specific search channels could explain diverse findings in the literature and, second, whether the type of the social contact matters for the starting wage. We answer these questions by using a unique combination of variables contained in the German Socio-Economic Panel (SOEP). SOEP is a large-scale household survey which is representative on the national level. It combines information about the type of the social contact providing a referral, distinguishing between family ties, friends and former colleagues, with rich information on the socio-economic and demographic background of the respondent as well as firm and occupation-specific characteristics. Most important, it includes information about the results of a short test of cognitive abilities, called the symbol-digit test (SDT), as well as personality traits3, which allows us to shed light on the role of observed and unobserved worker heterogeneity in selecting into specific search channels. In addition, we exploit a recent linkage of SOEP with the INKAR database4 giving us a possibility to account for county-level regional economic indicators. Our empirical analysis is guided by the theoretical framework developed in Stupnytska and Zaharieva (2015). This study considers a continuum of workers heterogenous in ability/productivity and searching for jobs when unemployed. Moreover, there are three search channels: formal applications and two informal channels – through family and professional networks. The model illustrates a strong selection of workers across the three search channels with high ability workers mostly entering jobs via referrals from professional contacts, low ability workers relying on their families and workers in the middle of the ability distribution entering jobs in a formal way. This selection pattern explains wage premia associated with professional recommendations borrowing from the seminal approach by Montgomery (1991) and Granovetter (1995). At the same time, it 1Addison and Portugal (2002), Kugler (2003), Margolis and Simonnet (2003), Delattre and Sabatier (2007), Goos and Salomons (2007), Ponzo and Scoppa (2010), G¨urtzgen and Pohlan (2024) 2We review and discuss both groups of studies in the next section. 3SOEP includes information about 10 personality traits which can be aggregated to the “Big Five” scale: conscientiousness, extraversion, agreeableness, openness, neuroticism. See McCrae and Costa (1999), Uysal and Pohlmeier (2011) and Caliendo et al. (2016) for further details. 4Indicators and Maps of Spatial and Urban Development 1 predicts wage penalties associated with referrals from family or close friends if worker abilities are not observable. We test these predictions by using the approaches based on inverse-probability weighting (IPW). These are characterized by having two stages. In the first stage we model the self-selection of workers into search channels with a (multinomial) logistic regression. The multinomial specification is used to account for the four search channels (family/friends/colleagues/formal) and generates propensity scores which are used to construct the weights in the estimation of the outcome variable. In the second stage we apply three weighting methods: pure weighted mean comparison (IPW), the augmented inverse probability weighting (AIPW) and the IPW - regression adjustment (IPWRA). All the approaches make the groups of workers, finding jobs though different search channels, comparable to each other in terms of observable characteristics. However, the latter two also model the outcome, which gives rise to the “doubly-robust” estimators (Imbens and Wooldridge 2009; Glynn and Quinn 2010;Kurz2022)5. First, our descriptive results reveal that there is a strong self-selection of workers based on formal qualification and working time, contributing together about 1/3 to the row wage penalty associated with referral hiring. Another 1/5 of the penalty stems from differences in the firm size and the likelihood of occupational mismatch. These findings support the ideas in Galenianos (2013) and Rebien et al. (2020) that referrals often lead to jobs in small firms. The observation that referred candidates are more likely to report occupational mismatch is in line with the evidence in Bentolila et al. (2010) putting forward the idea that referral hiring generates occupational mismatch between the initial qualification of the worker and the job requirements. Overall, we conclude that differences in observable characteristics contribute about 2/3 to the referral wage penalty. The IPW methods reveal a remaining wage penalty of −10%, which is persistent over time. Second, separating by the type of the social contact, we find that referrals from friends are associated with a significant wage penalty equal to −7%. At the same time, referrals from former colleagues show a robust wage premium between 9% and 10%. In order to study the role of unobserved heterogeneity behind these findings we compared the performance of respondents in the symbol digit text (SDT)6. This test was already used in other studies as a proxy for cognitive abilities (e.g. Heineck and Anger (2010), Richter et al. (2017)), moreover, Lang et al. (2007) show that the SDT outcome is sufficiently 5Another advantage of the IPW approach, in comparison to the standard OLS, is that it does not require the linearity assumption and allows us to perform the balancing checks of covariates before and after propensity score matching. Moreover, Imbens and Wooldridge (2009) highlight that in case the normalized differences between observable characteristics are considerable, the unconfoundedness assumption of the standard OLS might not be fulfilled properly, with results being extremely sensitive to the specification. Identification relying on the estimated propensity scores allows to avoid the issue. 6Every participant was provided with a mapping between pairs of symbols and digits on the screen. The participants had to match as many symbols to the corresponding numbers as they could. The number of correct, total and wrong answers was recorded after 30, 60 and 90 seconds. 2 correlated with test scores from more comprehensive intelligence tests. We find that workers referred by their colleagues perform better in the test than any other group. The second best group in terms of cognitive scores are those who found a job in a formal way, followed by those relying on the help of their families and friends. These results indicate substantial differences in cognitive abilities between the groups and support the sorting model in Stupnytska and Zaharieva (2015). In addition, we compared the noncognitive abilities based on the “Big Five” personality traits. We do not find systematic differences in non-cognitive abilities for workers using different search channels. One minor exception is the ability to forgive, which is highest among individuals referred by their family members. Third, we perform the heterogeneity analysis and separate the sample by gender and region. We find that the overall wage penalty from referrals is persistent for men and women and has a similar size. However, the finding that wage penalties/premia are associated with referrals from friends/colleagues is largely driven by the subsample of women. This is inline with the observation that the spread in cognitive scores between those finding the job with a help of friends/colleagues is larger for women in comparison to men. This provides another indirect support for the important role of cognitive abilities in explaining referral wage gaps. With respect to regional heterogeneity our results demonstrate that higher unemployment is associated with a stronger use of social networks. On the contrary, there is less reliance on networks in more densely populated regions. At the same time, wage penalties from referrals by friends and premia from referrals by colleagues are more pronounced in Eastern Germany. The rest of the paper is structured as follows. Section 2characterizes the current state of research and briefly summarizes the underlying theoretical model. Section 3describes the data and presents descriptive statistics. It is followed by section 4, which outlines the empirical methods. Next, section 5presents the estimation results for the overall sample, followed by the heterogeneity analysis by gender and West/East Germany in section 6. The robustness of the results is tested in section 7, while section 8concludes. 2 Literature Review and Theory 2.1 Related literature The literature on the effect of referrals on wages is rich but inconclusive. Whereas some studies find a significant positive link, others report wage penalties associated with entering the job via a referral (Topa 2011). Pellizzari (2010) highlights the puzzle by showing empirically that in the European Union “... premia and penalties to finding jobs through personal contacts are equally frequent and are of about the same size” (p. 494). Thus, in the following, we review sequentially the literature on wage premiums and wage penalties 3 associated with referrals paying particular attention to the underlying mechanisms and formulating several testable hypotheses to guide our empirical analysis. Wage premia: The seminal theoretical contributions date back to the work by Montgomery (1991) and Granovetter (1995). Montgomery (1991) shows that employee referrals help firms in screening the unobserved abilities of their applicants. In particular, Montgomery (1991) develops the idea of homophily by ability (or the inbreeding bias), meaning that high-ability workers are likely to have high-ability friends. Therefore, firms expect this relationship and offer higher wages to referred applicants. Hensvik and Skans (2016) confirm this result empirically and show that in Sweden entrants are more likely to be connected to high-ability employees than to low-ability ones (defined from the test scores or wages). Moreover, in their study, entering workers receive higher initial wages if they have a link to the incumbent employee. Another explanation for wage premia associated with referrals is presented by Simon and Warner (1992) and further extended by Galenianos (2014) and Dustmann et al. (2016). Simon and Warner (1992) show that referrals from employees reduce uncertainty about applicants’ productivity, which leads to higher reservation wages, higher starting wages and lower wage growth after the first period. All three studies provide empirical support for this mechanism based on the US, British and German data respectively, though the evidence in Dustmann et al. (2016) is only indirect (based on ethnic networks) and does not include detailed information about the type of social contact who provided a referral. A recent study by G¨urtzgen and Pohlan (2024) extends the model of employerlearning by adding formal screening activities. The authors use data from an employer survey for Germany to test their model and find a 1% wage premium. The result, however, is driven solely by the male subsample. One further study supporting the idea of wage premia is Kugler (2003). In this model, referrals reduce the cost of monitoring for firms, because workers exert peer pressure on co-workers, and make it cheaper to pay efficiency wages. Also Ioannides and Soetevent (2006) and Fontaine (2008) show theoretically that workers with a larger social network receive higher wages, whereas empirically wage premia are supported by Margolis and Simonnet (2003) for France and Goos and Salomons (2007)fortheUK. Wage penalties: Next, we consider the literature reporting wage penalties associated with referral hiring. Empirically wage penalties are documented by Pistaferri (1999) for Italy, Addison and Portugal (2002) for Portugal and Delattre and Sabatier (2007) for France. Ponzo and Scoppa (2010) argue that firms may offer low wages and hire low-ability family ties in the absence of more talented applicants. This is the idea of favouritism leading to wage penalties associated with referral hiring. Bentolila et al. (2010) develop a model showing that social contacts may generate a mismatch between the qualification of the worker and this worker’s productive advantage. Bentolila et al. (2010) present empirical evidence supporting this view for the US and European Union, 4 though the sample is limited to young workers (below 35 years of age). Horv´ath (2014)and Zaharieva (2018) extend this setup by introducing a network homophily parameter, i.e., the degree to which people form social connections with others from the same occupation. Galenianos (2013) and Rebien et al. (2020) put forward the idea that larger firms have more financial resources to advertise their vacancies compared to small firms, therefore, they receive many applications via the formal search channel making referral hiring redundant. Both studies provide empirical support for this idea. Premia/penalties depending on the contact type: Labini (2005), Stupnytska and Zaharieva (2015) and more recently Lester et al. (2021) emphasize the point that the type of social contact may play a crucial role for the sign of the referral wage gap. In particular, they show that referrals from professional contacts and former colleagues (weak ties) lead to a wage premium, whereas referrals from relatives and close friends (strong ties) generate a wage penalty. The underlying mechanism is based on the sorting of workers with different (unobserved) abilities across search channels. According to Lester et al. (2021) the sorting framework is supported in the US data, however, their dataset does not include a direct measure of worker cognitive abilities. Only a few other studies addressed this issue empirically, including Antoninis (2006) for Egypt, Meliciani and Radicchia (2011) for Italy and Cappellari and Tatsiramos (2015) for the UK. Even though their results are supportive of the described mechanism, the evidence in the former studies is descriptive, whereas the data in Cappellari and Tatsiramos (2015) does not include information about the type of social contact who provided a referral. Hence, in this paper we test the theoretical predictions from the sorting model by using German data (SOEP), which includes information about worker’s cognitive abilities and personality traits. We give a short overview of the sorting mechanism developed in Stupnytska and Zaharieva (2015) and summarize our hypotheses in the next section. Heterogeneity by gender and region: Further, we acknowledge that the referral wage gap could differ between the two gender groups and across locations. For example, Huffman and Torres (2002) and Wrzus et al. (2013) show empirically for the USA that the quality of referred jobs provided to a contact differs for men and women. Marmaros and Sacerdote (2002) study the effects of the peer networks and also report differences by gender and race. Zhou (2019) emphasizes that not only the size of the network is vital for a successful job search but also the willingness of the contacts to provide a referral. Germany presents another social context as a result of previous historical events. Despite the observable convergence in many factors, West Germany and East Germany are still different on such indicators as income levels, the unemployment rate as well as workrelated attitudes and gender roles (Schnabel 2016; Dirksmeier 2015). The latter can be highly-influential with regard to the quality of the job transmitted via referrals to men and women. Therefore, in the following analysis we go beyond the overall effect of referrals on wages and study separate effects by gender and region. 5 2.2 Theoretical considerations In this section we briefly describe the economic mechanism leading to wage premia or wage penalties from referrals in the model by Stupnytska and Zaharieva (2015). We test it later in the main body of the paper. Consider workers with heterogeneous productivities yi, i=1..p,wherepis the number of distinct worker groups. Differences in the productivity reflect differences in worker’s abilities, so that higher yiis associated with a higher wage. Workers’ productivity/ability is observed by the employer upon the match (in the course of the job interview), but it is not observable to the econometrician in the data. There are three search channels in the labour market. First, unemployed workers can find a job by sending applications to open vacancies, this is the formal channel of job search with a job-finding rate vsi. This rate is endogenous and depends on the search effort of the worker siand the stock of vacancies v. The formal channel is costly for workers in terms of effort with a cost function C(s)=s2/c. In addition, workers can find a job via their professional contacts at rate λi(weak ties) or family members λ0(strong ties). This setup gives rise to the following Bellman equation for unemployed workers: rUi=b+(λi+λ0)Ri+vmax ssRi−s2/c where bdenotes the unemployment benefit and Riis a worker job rent increasing in the worker type yidue to a higher wage. Maximizing the present value of unemployment Ui with respect to sleads to the result that the formal search effort of workers is increasing in the worker type yi(via Ri) but decreasing in the job-finding rate via professional contacts λi. Intuitively, this shows a disincentive effect: workers with a higher probability of getting a professional referral (λi) reduce their formal search effort si. Next, the authors derive λiby following the idea in Montgomery (1991) and assuming that workers form professional connections within their ability group. The idea of homophily by ability is also supported by Ioannides and Loury (2004). In general, homophily refers to the fact that people are more prone to maintain relationships with others who are similar to them. There can be homophily by race, gender, religion, skill or ability and it is generally a robust observation in social networks (see Jackson (2008) for an overview of research on network homophily). If the network size of every worker is fixed to n, the probability of finding a job via a professional referral is given by: λi=vai 1−μi μi [1 −(1 −μi)n] where aiis the advertising effort of firms and μiis the unemployment rate of workers in the productivity group i. In the above equation the term [1 −(1 −μi)n] stands for the probability that an incumbent type iemployee addressed by the firm has at least one unemployed friend who will take the job. In addition, this equation shows that firms can 6 referred for jobs differs significantly from the comparison group along multiple observable characteristics pointing out to the self-selection bias in the sample. Hence, further we discuss our estimation strategy building on the inverse probability weighting approach. We rely on the following family of methods, in contrast to the usual OLS estimator, as it allows not only to perform the balancing checks of selection-driving covariates and assure a common support, but it also does not require a linearity assumption in the weighting procedure (Glynn and Quinn 2010;Kurz2022). 4 Methods There are several challenges for performing the estimation. First, the self-selection bias is present both in the setup with aggregated and disaggregated networks. Second, the size of the treated and comparison groups is considerably different. As we can see from Table 3there are almost two times more respondents, who found a job with the help of networks, compared to those who used a formal individual application. Last but not least, is the disaggregated network setup, which calls for a simultaneous comparison of more than two search channels. To tackle these concerns we apply the inverse probability weighting techniques. These methods include the inverse probability weighting (IPW), the augmented inverse probability weighting (AIPW) and the inverse-probability weighting - regression adjustment (IPWRA). All of these methods have two stages. In the first stage we model the selfselection into search channels with a logistic regression. It is a binary logit in the aggregated setup with two search channels (formal/networks) and a multinomial logit in the disaggregated setup (formal/family/friends/colleagues). The corresponding propensity scores p(x)iare used as weights in the estimation of the outcome variable (starting wages) in the second stage. When estimating the average treatment effect under the IPW procedure, the weighted outcomes for both groups are calculated in the following way: ln(Ywti)=ln(Yti) p(x)i ln(Ywci)= ln(Yci) 1−p(x)i ,(1) where p(x)iis a propensity score estimated for individual iusing a (multinomial) logit approach. Ywti is a weighted outcome for observation iassigned to the treatment group t, while Ycti is a weighted outcome for individual iin the comparison group c.Boththe treatment and the comparison group weights are normalized within the respective group. The treatment effect is calculated as a simple difference of weighted means:  ATEIPW =Nt i=1 ln(Ywti) Nt+Nc −Nc i=1 ln(Ywci) Nt+Nc ,(2) where Ntand Nccorrespond to the number of observations in the treatment and the 13 comparison groups respectively. AIPW and IPWRA are extensions of the IPW and belong to the class of the “doubly robust” estimators: in both cases, self-selection and outcome are modeled, but for the consistency of the estimated coefficients, it is enough to specify at least one of the models correctly (Glynn and Quinn 2010;Kurz2022). The main difference between the two methods lies in the second stage when the inverse probability weights are applied. The AIPW procedure estimates separate regression models of the outcome for each treatment group and obtains the treatment-specific predicted outcomes. It then computes the weighted means of the obtained values. In contrast, the IPWRA procedure fits weighted regression models of the outcome for each treatment group and estimates the treatmentspecific predicted outcomes. It then computes the difference. The baseline model for AIPW/IPWRA in the aggregated setup can be written in the following way: ln(Yi)=β0+β1D+X iβ2+Z rβ3+τt+i,(3) where Yiis a dependent variable, e.g., starting wages in the new job; Dis the treatment indicator, which equals to 1 if a job was found using networks; Xiis a vector of control variables on individual level; Zris a set of regional indicators, τtis a time-fixed effect and iis an error term. In the disaggregated setup with multiple referral types (family, friends and colleagues) we adjust the regression equation to account for the three different treatment arms. The equation becomes: ln(Yi)=β0+β1D1+β2D2+β3D3+X iβ4+Z rβ5+τt+i,(4) where Yiis again the dependent variable, while D1to D3are dummy variables corresponding to the three referral channels: friends, family or colleagues. The interpretation of other variables remains the same. The set of control variables Xiin the AIPW and the IPWRA outcome models includes sex, age, age squared, years of education, occupational mismatch/being in studies/no training, a dummy variable for migration background; a dummy for a long distance move as well as moving to another NUTS3 region in the previous year; change in a labour market status from non-participation to employment or job-to-job transition; being in a full-time employment (compared to a part-time arrangement), logarithm of total hours worked during the year; a dummy variable for being employed in construction and if the currently employing firm is of a middle size. The set of regional characteristics Zrincludes the share of migrants, the unemployment rate and the logarithm of the population density. In order to correct for the sample selection in the first stage of the estimation, we include the following predictors: age, sex, years of education, marital status, occupational mismatch, relocation to another county, living in a rural or urban region, being 14 employed full-time, yearly working hours are greater than the average among the working individuals in a given year, changes in the labour market status as described in the outcome model, working in a small, middle or large firm, migration background, number of children in the household. The set of regional characteristics Zris the same in both model stages. Moreover, we also include a dummy for a respondent living in the West or East Germany at the time of the survey. An important condition for the validity of the IPW techniques is a balance of covariates in the adjusted sample. We test if the covariates are balanced by using the χ2 test statistic proposed by Imai and Ratkovic (2014). This is an adaptation of GMMoveridentification test for the covariates balance. Since all the estimation methods have the same first stage, it is enough to calculate the p-values of the χ2-test for the IPW alone. We find that the null-hypothesis of covariates being balanced cannot be rejected (p-value=0.46)9. In addition to the IPW approach, we use propensity score matching (PSM) as a robustness check in the aggregated analysis (formal/networks)10. Moreover, we report the PSM results in the second stage based on a “pure” weighted-mean comparison as well as regression-based. Note that PSM can not be applied in the disaggregated analysis with multiple treatment groups. The matching procedure was stratified between the East and the West and we used a 1:1 calliper matching algorithm with replacement11.The selected calliper equals 0.15 of the SD of the propensity score. The propensity score distribution is reported in Figure 8in Appendix 9.3. The same matching vector is used as in the treatment assignments models in the PS weighing approach. After the PSM, the median bias is reduced from 10.4% to 0.9%, the Rubin’s B parameter equals to 7.9, and the variance ratio R is 1.012. We used the same set of control variables described above in the first stage of the PSM. However, since the first-stage model does not have to be saturated, the controls in the outcome stage may be extended with additional regressors (Guo and Fraser 2014). 9The tests were also performed for the subgroups, and the results are reported in Section 6.Withthe same balancing vector described above, for all of the subgroups the null-hypothesis cannot be rejected. Thus, the covariates are balanced. 10It is important to acknowledge that in the recent literature there are debates regarding the expected bias which might be introduced with the overcorrection. The paper by King and Nielsen (2019) criticized the application of PSM due to the possibility of introducing a higher bias compared to the unmatched case. The authors proposing resorting to the Mahalanobis distance or entropy matching instead. However, this paper sparkled debate caused by the nature of the simulated data used. The simulation study by Vable et al. (2019) covered different types of data and highlighted that a proper common support was essential for the matching methods to be unbiased and superior to OLS in case of dissimilar estimates. Moreover, unlike OLS, PSM allows to tackle the self-selection bias (Titus 2007). 11Austin (2014) compared 12 different matching approaches and found that calliper matching has performance of at least as good as other algorithms and superior in terms of bias reduction. 12Rubin’s B is an absolute standardized difference in means of the propensity score between treated and comparison groups. Rubin’s R is the ratio of variances in the propensity scores of the treated and comparison groups. In order to consider sample being balanced, the first parameter should be below 25%, while the second one is [0.5,2] (Rubin 2001). 15 Consequently, the set of additional regressors includes such variables as age squared, tenure, total years spent in unemployment and household size. In order to account for occupational differences we add a set of dummy variables for being in a particular occupation (following the NACE rev. 1.1 classification). Moreover, we extend the set of regional indicators with the federal state fixed effects and a binary indicator for the region being rural or urban. The exact construction of the main variables is discussed in detail in Appendix 9.2. 5Results In this section we apply the methodology discussed above. We start by considering the aggregated setup with a binary search channel indicator (formal vs. network search). Our results are summarized in Table 4where columns (1)-(3) correspond to the IPW while columns (4)-(5) are based on the PSM approach. The overall effect of finding a position via networks on the logarithm of starting wages is negative and robust: it is statistically significant according to all estimation methods. Although the estimated wage penalty associated with referrals varies between −8% and −10% a linear hypothesis about the intra-model equality of the estimated coefficients cannot be rejected13.Moreover,Table13 in Appendix 9.1 shows that referrals from friends and relatives make 84% of all referral cases with the recommendations from colleagues making up the remaining 16%. This means that referrals from strong ties are much more frequent in the data than professional recommendations. Hence, we find that hypothesis H1 is strongly supported by the data. In the second step we perform the disaggregated analysis by distinguishing between the three types of network ties: friends, relatives, and colleagues. Our results are summarized in columns (6)-(8) of Table 4. It shows that the coefficients for relatives and friends are negative. Referrals from friends are associated with a significant wage penalty equal to −7%. At the same time weak ties show a robust positive effect on starting wages between 9% and 10%. The observed wage differences are in line with the theoretical model and support hypotheses H2 and H3. However, at this step the empirical support for these hypotheses is still incomplete since the underlying mechanism in the model is based on ability/productivity differences of worker groups using different search channels. Therefore, in the third step we use a proxy for the cognitive abilities of workers and compare them across the channels. In the years 2006, 2012 and 2016 SOEP has introduced a short numerical test to evaluate cognitive abilities of the respondents. This test is referred to as a symbol-digit 13As for the sample sizes of the estimated groups, the PSM sample is smaller in terms of the number of the unique observations but the frequency weights compensate for this difference. The IPW estimates do not result in the sample shrinkage but are restricted by the collected information on control variables in the treatment assignment and/or outcome models. 16 Table 4: Effects of the job search channels compared to the formal approach Networks: overall Networks: disaggregated (1) (2) (3) (4) (5) (6) (7) (8) IPW AIPW IPWRA PSM PSM IPW AIPW IPWRA Networks -0.074*** -0.106*** -0.104*** -0.131*** -0.082*** (0.013) (0.018) (0.018) (0.025) (0.023) Networks: friends -0.071*** -0.071*** -0.072*** (0.018) (0.021) (0.021) Networks: relatives -0.023 -0.001 -0.006 (0.039) (0.035) (0.035) Networks: colleagues 0.063 0.087** 0.101*** (0.041) (0.040) (0.039) Observations 10133 6126 6126 12677 7529 5986 5100 5100 Treatment assignment model logit logit logit logit logit mlogit mlogit mlogit Outcome model Weig. Avg. linear by ML linear Weig. Avg linear Weig. Avg. linear by ML linear Control variables in outcome model No Yes Yes No Yes No Yes Yes Notes: *** p<0.01, ** p<0.05, * p<0.10; Standard errors are clustered at NUTS3 level. Control variables include demographic controls on a personal/hh level as well as regional level control variables. The outcome variable is logarithm of the gross monthly wages. test (SDT) and reveals innate abilities of the individual and the speed of solving new tasks. For a detailed description see Lang et al. (2007) who show that SDT outcomes are sufficiently correlated with test scores of a more comprehensive intelligence test. During the course of the test participants are given 1.5 minutes (90 seconds) to connect the symbols with the digits, given a pre-defined conversion table. For instance, digit “1” corresponds to a symbol “*” and digit “2” to “?”. The question which the respondents are required to answer is “Which number corresponds to the symbol “*””? The number of correct, false and total answers is measured after 30, 60 and 90 seconds in the test. This test was already used as a proxy for the in-born cognitive abilities in other studies (Heineck and Anger 2010; Richter et al. 2017). Figure 2shows the distribution of the number of correct answers after 30, 60 and 90 seconds, aggregated over all available test years. The mean score is increasing over time from 9.4 correct answers in 30 seconds to 29.7 in 90 seconds. Moreover, the variance of the distribution increased as well, almost doubling from 5.74 to 12.52. Nevertheless, in all three cases approximately 5% of the respondents haven’t given any correct answers. Unfortunately, the available sample size is low, especially in the disaggregated network setup starting in 2014. Consequently, assuming that the test scores of adults are stable over several years, we extrapolate the results of the test performed in 2016 to the other years of interest in order to achieve a larger sample size14. Table 5reports the distribution of imputed test scores by search channels. First three lines of the table show a number of correct answers in 30, 60 and 90 seconds respectively. 14Raw gaps in cognitive abilities before extrapolation are provided in table 17 in appendix 9.3.However, even after the extrapolation the sample size is not sufficiently large for the cognitive scores to be included in the IPW regressions, so we are bound to the descriptive analysis of test scores. 17 The next three lines are dedicated to the number of wrong answers in the same time period, and the last lines demonstrate the total number of answers given. The groups have several profound statistically significant differences. Those who found a position with the help of colleagues, gave more correct answers and made fewer mistakes than any other group. The second best group in terms of cognitive scores are those who found a job individually. They make significantly more mistakes than those who located a job with the help of colleagues. At the same time, they gave more correct and total answers and made fewer mistakes than those who found a job with the help of friends or family. These discrepancies reveal that the average cognitive ability indeed differs between the groups, with relatively low-ability respondents relying on their friends or family to find a job, middle-ability ones finding a job individually and high-ability respondents acquiring a position via referrals from colleagues. We conclude that this evidence provides further support for hypotheses H2 and H3. Figure 2: Distributions of scores in the Symbol digit test at different durations Further, we find that differences in the SDT scores are robust to controlling for the years of education, mode of the interview15, and ID of an interviewer, despite a further reduction in the sample size. For instance, the number of wrong answers given in 60 seconds remains statistically significantly different between the groups. More detail is available in Table 15 in Appendix 9.3. Another vital part of individual characteristics includes non-cognitive abilities. For example, stronger self-confidence or better negotiation skills could have an affect on the 15This is the way the questionnaire was administrated, e.g., a computer assisted personal interview (CAPI) or a paper assisted one (PAPI). 18 Table 5: Numeric tests (cognitive ability), disaggregated network types. (1) (2) (3) (4) T-test Individually NW: friends NW: family NW: colleagues Difference Variable Mean/SE Mean/SE Mean/SE Mean/SE (1)-(2) (1)-(3) (1)-(4) (2)-(3) (2)-(4) (3)-(4) Correct asw. (30 sec) 11.148 10.734 11.035 11.428 0.414* 0.113 -0.280 -0.301 -0.694** -0.393 (0.292) (0.176) (0.279) (0.246) Correct asw. (60 sec) 23.499 22.587 23.015 23.808 0.911** 0.484 -0.309 -0.427 -1.220* -0.793 (0.489) (0.318) (0.486) (0.482) Correct asw. (90 sec) 35.294 34.016 34.435 35.684 1.278** 0.859 -0.391 -0.419 -1.669* -1.249 (0.614) (0.476) (0.677) (0.657) Wrong asw. (30 sec) 0.211 0.255 0.257 0.166 -0.044 -0.046 0.045 -0.003 0.089** 0.091** (0.018) (0.019) (0.038) (0.025) Wrong asw. (60 sec) 0.361 0.490 0.496 0.260 -0.129*** -0.136** 0.101*** -0.007 0.230*** 0.237*** (0.017) (0.038) (0.046) (0.023) Wrong asw. (90 sec) 0.574 0.732 0.803 0.547 -0.158*** -0.229** 0.028 -0.071 0.186** 0.257** (0.039) (0.044) (0.089) (0.035) Total asw. (30 sec) 11.359 10.989 11.293 11.594 0.370* 0.066 -0.235 -0.304 -0.605* -0.301 (0.283) (0.179) (0.274) (0.234) Total asw. (60 sec) 23.859 23.077 23.511 24.067 0.782* 0.348 -0.208 -0.434 -0.990 -0.556 (0.492) (0.309) (0.452) (0.478) Total asw. (90 sec) 35.868 34.748 35.238 36.231 1.120** 0.630 -0.363 -0.490 -1.483* -0.993 (0.626) (0.467) (0.646) (0.672) Number of observations 1717 1945 442 484 3662 2159 2201 2387 2429 926 Number of clusters 16 16 16 16 16 16 16 16 16 16 Notes: Values are winsorised at 1 and 99 percentiles. Imputed for all years but 2016, assuming the results of the test being constant for the same respondent over 5 years. The value displayed for t-tests are the differences in the means across the groups. Standard errors are clustered at federal state level. Observations are weighted using variable phrf as aweight weights.***, **, and * indicate significance at the 1, 5, and 10 percent critical level. Table 17 based on original values is available in Appendix 9.3. starting wages. In order to test if these differences are relevant or not, we use a set of questions on the personality traits contained in SOEP (see McCrae and Costa (1999)for further details). These questions can be used to construct the “Big Five” personality traits as shown by Uysal and Pohlmeier (2011) and Caliendo et al. (2016). We use the following variables: being a thorough worker and carrying out tasks efficiently (proxies for conscientiousness), being original and valuing artistic experiences (proxies for the openness to experience), being sociable and communicative (proxies for extraversion), being able to forgive and friendly with others (proxies for agreeableness), worrying a lot and being nervous (proxies for neuroticism). However, Table 16 in Appendix 9.3 shows no systematic statistically significant differences between the groups. One minor exception is an ability to forgive, whereby more forgiving individuals are more often referred by their family members compared to those searching formally or recommended by friends. In the final step we study the implications of regional characteristics. We find that county-specific factors have pronounced influence on the self-selection of workers into formal job search versus networks. Table 6demonstrates the coefficients of the first-stage regressions for the regional variables, which have been included. In all cases - for the aggregated and disaggregated setup - higher regional unemployment rate is associated with a higher probability of utilizing networks. On the other hand, the logarithm of the population density in a county is negatively related to the probability of using a network tie for every group but colleagues. The contribution of the share of migrants is most 19 pronounced in the case of friendship ties. This is in line with the previous literature showing that immigrant workers often rely on their network of friends for obtaining jobs (see, for example, Dustmann et al. (2016)). Table 6: Effects of regional coefficients on self-selection in job search PSM IPW (1) NW: All (2) NW: Friends (3) NW: Family (4) NW: Colleagues Unemployment rate 0.049*** 0.045** 0.055** 0.043* (0.010) (0.019) (0.024) (0.025) Log of population density -0.131*** -0.225*** -0.303*** -0.008 (0.046) (0.077) (0.094) (0.097) Share of migrants 0.006 0.035** 0.005 0.002 (0.008) (0.014) (0.018) (0.018) Observations 10764 2363 564 490 Notes: *** p<0.01, ** p<0.05, * p<0.10; Standard errors are clustered at NUTS3 level. Control variables include demographic controls on a personal/hh level as well as regional level control variables. The outcome variable is a probability of belonging to a group based on the job search method. Logit is used to model the binary choice; multinomial logit is used to predict the results disaggregated by a network type. Number of observations is weighted in the PSM. In the IPW, number of observations presents available observations by group. The total number of observations in IPW is 5986. To sum up, when evaluating the network effect on starting wages in Germany, the coefficients are negative and robust indicating an approx. -9% lower starting wage. However, when the results are disaggregated by the network types, we can see that jobs found via friends and family are associated with a wage penalty of approximately -7% compared to the formal individual approach. Conversely, using the help of colleagues is associated with a 9% higher starting wage. In line with the theoretical model, the self-selection into job search methods is correlated with the results of the cognitive abilities tests: the group of those, who found a job with the help of colleagues shows the best outcomes. 6 Heterogeneity Analysis 6.1 Descriptive Analysis This section is dedicated to the analysis of regional and gender differences in the job search channels and associated wages. Men and women tend to have different behavioural patterns in labour markets. However, even if the behavior is similar market forces may generate different labour market outcomes (Perez 2019), thus, we study the implications of referral hiring separately for men and women. In addition, we account for the long 20 historical division of Germany between the East and the West. Figure 3presents a spatial distribution of job search via networks (blue) and via the formal individual approach (green). Note that the sample is restricted to only these two channels. Although there is no federal state, where the formal job search method is more prevalent than networking, the spatial patterns are present for both. Finding a job formally is more common in the south (e.g. the state of Baden-W¨urtemberg) and in the city-states Berlin and Hamburg, whereas the use of networks is most common in eastern federal states and in the west (Rheinland Pfalz). Figure 7in Appendix 9.1 demonstrates an overlap of regional differences in the job search methods with gender. The relative importance of the formal channel in the south is mostly driven by the subsample of men. Whereas the prevalence of the network channel in the west is driven by the subsample of women. We find that up to 61% of women in the west tend to find a position with the help of networks, while in the other parts of the country the share is lower. Figure 3: Job search channels in the years 2002-2019 In terms of the observable demographic characteristics of men and women, the most profound contrast is the following. Although men and women in the comparison group are statistically significantly different only in three characteristics16, the groups of men and women, who found their positions with the help of networks, are different in all 16The share of transitions from unemployment to employment is higher among men, while the share of respondents employed part-time and the average number of kids in a household are higher among women. The table with descriptive evidence by gender is available from the authors upon a request. 21 demographic characteristics but the share of mid-sized firms. This evidence supports our motivation to perform the analysis for subgroups distinguishing between the two broad regions (East vs. West) and the two gender groups. 6.2 Separating the effects by gender Table 7demonstrates the estimation results in the aggregated setup with a binary search channel indicator. The results for men are contained in Panel A, while the coefficients for women are reported in Panel B. Both panels indicate wage penalties associated with referral hiring, however, for men the results are not robust and vary from statistically insignificant to −9%. For women, on the other hand, all the estimators have overlapping confidence bounds and a mean close to −11% in starting wages. This suggests that the penalties from referral hiring are larger for women. Table 7: Effects of finding a job with the help of networks compared to the formal approach, sample splitting by sex (1) (2) (3) (4) (5) IPW AIPW IPWRA PSM PSM Panel A: Aggregated Networks, Men Networks -0.036* -0.096*** -0.088*** -0.077** -0.039 (0.019) (0.029) (0.029) (0.037) (0.034) Observations 4264 2168 2168 5569 2768 Panel B: Disaggregated Networks, Women Networks -0.105*** -0.107*** -0.105*** -0.161*** -0.091*** (0.018) (0.021) (0.021) (0.034) (0.029) Observations 5869 3958 3958 7108 4761 Treatment assignment model logit logit logit logit logit Outcome model Weig. Avg linear by ML linear Weig. Avg linear Control variables in outcome model No Yes Yes No Yes Notes: *** p<0.01, ** p<0.05, * p<0.10; Standard errors are clustered at NUTS3 level. Control variables include demographic controls on a personal/hh level as well as regional level controls and federal state dummies. The outcome variable is logarithm of the gross monthly wages. The disaggregated setup in Table 8highlights an even more heterogeneous picture for the two groups. For men, none of the coefficients in any of the methods is statistically significantly different from zero. Moreover, the estimated amplitude is approaching zero as well. For women, on the other hand, the coefficients associated with referrals from friends and colleagues are robust and statistically significant in all the three estimation methods. In line with the theoretical predictions, those, who found a job with the help of 22 Figure 6: Effects of finding a job via networks on wages in consequent years 8 Discussion and Conclusions In this paper we study the relationship between the job search channel and the starting wage. The analysis is focused on comparing the wages of referred with non-referred applicants by using the inverse-probability weighting approach. Our results show a stable negative effect of about -10% associated with finding a job with the help of a social contact in Germany. This wage gap is robust if estimated by different methods and using various subsamples, moreover, it remains stable several years after the start of the employment relationship. Disaggregating the effect by the type of the social contact, two opposite patterns are revealed - while jobs found via friends result in -7% lower starting wages, the positions found with the help of colleagues are associated with a wage premia of 9%. Combining this analysis with information on the cognitive ability score from the symbol digit test (SDT) we find that applicants recommended by their colleagues perform better in the ability test than those finding jobs in the formal way. In contrast, applicants finding jobs with a help of a friend perform worse in the test than any other worker group. These findings are consistent with the sorting model of job search as developed by Stupnytska and Zaharieva (2015) and Lester et al. (2021) as well as the seminal work by Montgomery (1991) and Granovetter (1995). The heterogeneity analysis shows significant differences by gender: for men none of the coefficients is statistically significant, while for women the differences are more pronounced. It is consistent with a large variation in cognitive scores documented for women and indicating a stronger sorting of women across search channels compared to men. Concluding the paper, we discuss the limitations of our analysis. The first limitation is in the nature of the data since SOEP has limited information about the employer, thus, unobserved employer heterogeneity is not accounted for. Further, our study does 29 not cover the respondents who found a job with the help of an employment agency or returned to a previous employer. Another restriction is self-selection in the labour market participation, which results in wages of non-participants not being observed. On the one hand, different methods allow controlling for this part of self-selection. On the other hand, the usual assumption is that the sign of self-selection is known. Machado (2017) argues that for women in the labour market it is not the case, with self-selection being positive for the lower deciles of the ability distribution and negative for the higher ones. Thus, one cannot claim the exact direction of the bias. However, if the usual positive self-selection is assumed, this paper estimates the lower bound of the effect. References Addison, John T and Pedro Portugal (2002). “Job search methods and outcomes”. In: Oxford Economic Papers 54.3, pp. 505–533. Antoninis, Manos (2006). “The wage effects from the use of personal contacts as hiring channels”. In: Journal of Economic Behavior & Organization 59.1, pp. 133–146. Austin, Peter C (2014). “A comparison of 12 algorithms for matching on the propensity score”. In: Statistics in medicine 33.6, pp. 1057–1069. Bentolila, Samuel, Claudio Michelacci, and Javier Suarez (2010). “Social Contacts and Occupational Choice”. In: Economica 77.305, pp. 20–45. issn: 1468-0335. Caliendo, Marco, Steffen K¨unn, and Martin Weissenberger (2016). “Personality traits and the evaluation of start-up subsidies”. In: European Economic Review 86, pp. 87–108. Cappellari, Lorenzo and Konstantinos Tatsiramos (2015). “With a Little Help from My Friends? Quality of Social Networks, Job Finding and Job Match Quality § ”. In: European Economic Review 78, pp. 55–75. Delattre, Eric and Mareva Sabatier (2007). “Social capital and wages: an econometric evaluation of social networking’s effects”. In: Labour 21.2, pp. 209–236. Dirksmeier, Peter (2015). “The intricate geographies of gender ideologies in Germany”. In: Geoforum 64, pp. 12–24. Dustmann, Christian et al. (2016). “Referral-based job search networks”. In: The Review of Economic Studies 83.2, pp. 514–546. Fontaine, Francois (2008). “Why are similar workers paid differently? The role of social networks”. In: Journal of Economic Dynamics and Control 32.12, pp. 3960–3977. Galenianos, Manolis (2013). “Learning about match quality and the use of referrals”. In: Review of Economic Dynamics 16.4, pp. 668–690. — (2014). “Hiring through referrals”. In: Journal of Economic Theory 152, pp. 304–323. Glynn, Adam N and Kevin M Quinn (2010). “An introduction to the augmented inverse propensity weighted estimator”. In: Political analysis 18.1, pp. 36–56. Goebel, Jan et al. (2018). “The German Socio-Economic Panel (SOEP)”. In: Jahrb¨ucher f¨ur National¨okonomie und Statistik/ Journal of Economics and Statistics. Goos, Maarten and Anna Salomons (2007). “Dangerous liaisons: a social network model for the gender wage gap”. In: Available at SSRN 1021699. Granovetter, Mark (1995). Getting a Job: A Study of Contacts and Careers. University of Chicago Press. Granovetter, Mark S (1973). “The strength of weak ties”. In: American journal of sociology 78.6, pp. 1360–1380. 30 Guo, Shenyang and Mark W Fraser (2014). Propensity score analysis: Statistical methods and applications. Vol. 11. SAGE publications. G¨urtzgen, Nicole and Laura Pohlan (2024). “Do employers learn more from referrals than from other recruitment channels?” In: Labour Economics, p. 102574. Heineck, Guido and Silke Anger (2010). “The returns to cognitive abilities and personality traits in Germany”. In: Labour economics 17.3, pp. 535–546. Hensvik, Lena and Oskar Nordstr¨om Skans (2016). “Social networks, employee selection, and labor market outcomes”. In: Journal of Labor Economics 34.4, pp. 825–867. Horv´ath, Gergely (2014). “Occupational mismatch and social networks”. In: Journal of Economic Behavior & Organization 106, pp. 442–468. Huffman, Matt and Lisa Torres (2002). “It’s not only “who you know” that matters: Gender, personal contacts, and job lead quality”. In: Gender & Society 16.6, pp. 793– 813. Imai, K. and M. Ratkovic (2014). “Covariate balancing propensity score”. In: Journal of the Royal Statistical Society Series B: Statistical Methodology 76.1, pp. 243–263. Imbens, Guido W and Jeffrey M Wooldridge (2009). “Recent developments in the econometrics of program evaluation”. In: Journal of economic literature 47.1, pp. 5–86. Ioannides, Yannis M and Linda Datcher Loury (2004). “Job Information Networks, Neighborhood Effects, and Inequality”. In: Journal of Economic Literature 42.4, pp. 1056– 1093. Ioannides, Yannis M and Adriaan R Soetevent (2006). “Wages and employment in a random social network with arbitrary degree distribution”. In: American Economic Review 96.2, pp. 270–274. Jackson, Matthew O. (2008). Social and Economic Networks. Princeton University Press. King, Gary and Richard Nielsen (2019). “Why propensity scores should not be used for matching”. In: Political analysis 27.4, pp. 435–454. Kugler, Adriana D (2003). “Employee referrals and efficiency wages”. In: Labour economics 10.5, pp. 531–556. Kurz, Christoph F (2022). “Augmented inverse probability weighting and the double robustness property”. In: Medical Decision Making 42.2, pp. 156–167. Labini, Mauro Sylos (2005). Social Networks and Wages: It is all about connections! Citeseer. Lang, Frieder R et al. (2007). “Assessing Cognitive Capacities in Computer-Assisted Survey Research: Two Ultra-Short Tests of lntellectual Ability in the German SocioEconomic Panel (SOEP)”. In: Journal of Contextual Economics–Schmollers Jahrbuch 1, pp. 183–191. Lester, Benjamin R., David Rivers, and Giorgio Topa (2021). “The heterogeneous impact of referrals on labor market outcomes”. In: FRB of New York Staff Report 987. Machado, Cecilia (2017). “Unobserved selection heterogeneity and the gender wage gap”. In: Journal of Applied Econometrics 32.7, pp. 1348–1366. Margolis, David N and V´eronique Simonnet (2003). “Educational track, networks and labor market outcomes”. In: Networks and Labor Market Outcomes (January 2003). Marmaros, David and Bruce Sacerdote (2002). “Peer and social networks in job search”. In: European Economic Review 46.4-5, pp. 870–879. issn: 0014-2921. McCrae, Robert R. and Paul T. Costa (1999). “A five-factor theory of personality”. In: In: Pervin, L.A., John, O.P. (Eds.), Handbook of Personality: Theory and Research, pp. 139–153. 31 Meliciani, Valentina and Debora Radicchia (2011). “The informal recruitment channel and the quality of job-worker matches: an analysis on Italian survey data”. In: Industrial and corporate change 20.2, pp. 511–554. Montgomery, James D (1991). “Social Networks and Labor-Market Outcomes: Toward an Economic Analysis”. In: 81.5, pp. 1408–1418. Pellizzari, Michele (2010). “Do Friends and Relatives Really Help in Getting a Good Job?:” in: IRL Review 63.3, pp. 494–510. issn: 00197939. Perez, Caroline Criado (2019). Invisible women: Data bias in a world designed for men. Abrams. Pistaferri, Luigi (1999). “Informal networks in the Italian labor market”. In: Giornale degli economisti e annali di economia, pp. 355–375. Ponzo, Michela and Vincenzo Scoppa (2010). “The use of informal networks in Italy: Efficiency or favoritism?” In: The Journal of Socio-Economics 39.1, pp. 89–99. Rebien, Martina, Michael Stops, and Anna Zaharieva (2020). “Formal search and referrals from a firm’s perspective”. In: International Economic Review 61.4, pp. 1679–1748. Richter, David et al. (2017). SOEP scales manual (updated for SOEP-Core v32. 1).Tech. rep. SOEP survey papers. Rubin, Donald B (2001). “Using propensity scores to help design observational studies: application to the tobacco litigation”. In: Health Services and Outcomes Research Methodology 2, pp. 169–188. Schnabel, Claus (2016). “United, yet apart? A note on persistent labour market differences between western and eastern Germany”. In: Jahrb¨ucher f¨ur National¨okonomie und Statistik 236.2, pp. 157–179. Simon, Curtis J and John T Warner (1992). “Matchmaker, matchmaker: The effect of old boy networks on job match quality, earnings, and tenure”. In: Journal of labor economics 10.3, pp. 306–330. Sprengholz, Maximilian, Anna Wieber, and Elke Holst (2022). “Gender identity and wives’ labor market outcomes in West and East Germany between 1983 and 2016”. In: Socio-Economic Review 20.1, pp. 257–279. Stupnytska, Yuliia and Anna Zaharieva (2015). “Explaining U-shape of the referral hiring pattern in a search model with heterogeneous workers”. In: Journal of Economic Behavior & Organization 119, pp. 211–233. Titus, Marvin A (2007). “Detecting selection bias, using propensity score matching, and estimating treatment effects: An application to the private returns to a master’s degree”. In: Research in Higher Education 48, pp. 487–521. Topa, Giorgio (2011). “Labor markets and referrals”. In: Handbook of social economics. Vol. 1. Elsevier, pp. 1193–1221. Uysal, Derya and Winfried Pohlmeier (2011). “Unemployment duration and personality”. In: Journal of Economic Psychology 32.6, pp. 980–992. Vable, Anusha M et al. (2019). “Performance of matching methods as compared with unmatched ordinary least squares regression under constant effects”. In: American journal of epidemiology 188.7, pp. 1345–1354. Wrzus, Cornelia et al. (2013). “Social network changes and life events across the life span: a meta-analysis.” In: Psychological bulletin 139.1, p. 53. Zaharieva, Anna (2018). “On the optimal diversification of social networks in frictional labour markets with occupational mismatch”. In: Labour Economics 50, pp. 112–127. Zhou, Min (2019). “Gender Differences in the Provision of Job-Search Help”. In: Gender and Society 33.5, pp. 746–771. issn: 08912432. 32 9 Appendix 9.1 Detailed demographic characteristics Table 12: Blinder-Oaxaca decomposition Panel A: overall decomposition Estimated values Individual 7.149 (0.02)*** Networking: overall 6.832 (0.02)*** Total difference 0.316 (0.02)*** Endowments 0.210 (0.02)*** Coefficients 0.108 (0.02)*** Panel B: detailed disaggregation Endowments Coefficients Interaction Age -0.023 (0.02) -0.416 (0.49) 0.006 (0.01) Age, squared 0.029 (0.02)* 0.246 (0.24) -0.008 (0.01) Federal State 0.001 (0.00) 0.044 (0.04) -0.001 (0.00) Education: less than high school -0.005 (0.00) -0.005 (0.03) 0.001 (0.01) Education: more than high school 0.007 (0.01) -0.011 (0.02) -0.006 (0.01) Job-to-job transition 0.000 (0.00) 0.018 (0.02) 0.001 (0.00) Firm size by empl: No employees 0.000 (.) 0.000 (.) 0.000 (.) Firm size by empl: Small 0.005 (0.01) 0.068 (0.06) -0.020 (0.02) Firm size by empl: Middle 0.001 (0.00) 0.054 (0.04) 0.001 (0.00) Firm size by empl: Large 0.036 (0.01)*** 0.049 (0.05) 0.021 (0.02) First Job Full Time -0.001 (0.00) -0.049 (0.03)* 0.001 (0.00) # of children in HH 0.003 (0.00) 0.002 (0.04) -0.000 (0.00) Number of Persons in HH -0.003 (0.00) -0.017 (0.05) 0.001 (0.00) Annual Work Hours of Individual 0.056 (0.01)*** -0.088 (0.03)*** -0.010 (0.00)*** The youngest child is in kita age or smaller -0.000 (0.00) 0.015 (0.01) -0.001 (0.00) Long-term unemployement in county (INKAR) -0.000 (0.00) -0.018 (0.09) 0.000 (0.00) Married, living together -0.000 (0.00) -0.024 (0.02) 0.001 (0.00) Migration background 0.000 (0.00) 0.016 (0.01) -0.003 (0.00) NACE code (rev 1.1) 0.001 (0.00) 0.034 (0.04) -0.000 (0.00) Occupation & Training: mismatch 0.033 (0.01)*** -0.013 (0.02) 0.003 (0.00) Occupation & Training: in training 0.001 (0.01) 0.001 (0.01) -0.000 (0.00) Occupation & Training: no training 0.015 (0.00)*** -0.000 (0.01) 0.000 (0.00) Tenure -0.002 (0.00)*** 0.012 (0.01) -0.003 (0.00) Population-workplace density in county (INKAR) 0.005 (0.00)** -0.019 (0.01) -0.003 (0.00) Rural or urban region -0.000 (0.00) 0.008 (0.03) 0.000 (0.00) Moved to another county 0.005 (0.00)** 0.002 (0.00) 0.002 (0.00) Sex -0.006 (0.00)* -0.002 (0.01) 0.000 (0.00) Survey Year -0.002 (0.00) -19.524 (9.88)** 0.001 (0.00) Total # of years a resp. spent unempl from 2001 0.004 (0.00)* 0.025 (0.03) -0.001 (0.00) Transition from unempl. to empl. 0.002 (0.00) 0.007 (0.01) 0.000 (0.00) Unemployment rate in county (INKAR) 0.005 (0.00)* 0.015 (0.05) -0.001 (0.00) Years of education 0.045 (0.01)*** 0.215 (0.18) 0.015 (0.01) Constant 19.461 (9.86)** Notes: Standard errors are clustered at NUTS3 level and reported in paranthesis. *, ** and *** indicate 10-, 5and 1-percent critical level. The dependent variable is logarithm of gross monthly wages. 33 Table 13: Demographic characteristics conditional on the job search method, disaggregated network types. (1) (2) (3) (4) (2)-(1) (3)-(1) (4)-(1) Individually Networks: friends Networks: family Networks: colleagues Pairwise t-test Variable N/Clusters Mean/(SE) N/Clusters Mean/(SE) N/Clusters Mean/(SE) N/Clusters Mean/(SE) N/Clusters Mean difference N/Clusters Mean difference N/Clusters Mean difference Natural logarithm of gross wages 4269 7.442 4838 7.147 1220 7.018 1152 7.683 9107 -0.294*** 5489 -0.424*** 5421 0.242*** 96 (0.028) 95 (0.031) 91 (0.056) 88 (0.040) 96 96 96 Men 4305 0.435 4890 0.521 1236 0.484 1158 0.546 9195 0.086*** 5541 0.049 5463 0.111*** 96 (0.016) 95 (0.012) 91 (0.025) 88 (0.021) 96 96 96 Age, years 4305 35.272 4890 36.934 1236 34.252 1158 36.665 9195 1.661*** 5541 -1.020 5463 1.393*** 96 (0.262) 95 (0.257) 91 (0.634) 88 (0.420) 96 96 96 Years of educations 4081 12.894 4566 11.662 1145 11.539 1101 13.163 8647 -1.231*** 5226 -1.354*** 5182 0.270* 96 (0.067) 95 (0.092) 91 (0.125) 88 (0.131) 96 96 96 Married, living together 4289 0.370 4859 0.420 1231 0.435 1157 0.350 9148 0.050*** 5520 0.065** 5446 -0.020 96 (0.012) 95 (0.015) 91 (0.026) 88 (0.022) 96 96 96 # of HH members 4305 2.534 4890 2.695 1236 3.095 1158 2.429 9195 0.161*** 5541 0.561*** 5463 -0.105* 96 (0.039) 95 (0.040) 91 (0.072) 88 (0.054) 96 96 96 # of children in HH 4305 0.899 4890 1.091 1236 1.055 1158 0.829 9195 0.192*** 5541 0.156** 5463 -0.070 96 (0.030) 95 (0.025) 91 (0.056) 88 (0.038) 96 96 96 Migration background 4305 0.326 4890 0.459 1236 0.412 1158 0.273 9195 0.134*** 5541 0.087** 5463 -0.053** 96 (0.017) 95 (0.028) 91 (0.039) 88 (0.022) 96 96 96 Transition from unempl. to empl. 3566 0.105 3865 0.120 976 0.102 997 0.021 7431 0.015 4542 -0.003 4563 -0.084*** 96 (0.010) 95 (0.012) 90 (0.016) 87 (0.006) 96 96 96 Job-to-job transition 4305 0.587 4890 0.533 1236 0.474 1158 0.722 9195 -0.054*** 5541 -0.112*** 5463 0.135*** 96 (0.018) 95 (0.022) 91 (0.031) 88 (0.023) 96 96 96 Main job full-time 4008 0.728 4556 0.757 1095 0.703 1100 0.767 8564 0.029** 5103 -0.025 5108 0.039 96 (0.013) 95 (0.012) 90 (0.022) 88 (0.023) 96 96 96 Annual work hours 4305 1444.608 4890 1262.939 1236 1167.793 1158 1732.927 9195 -181.669*** 5541 -276.814*** 5463 288.319*** 96 (27.660) 95 (30.786) 91 (49.016) 88 (33.418) 96 96 96 Occupational mismatch 4264 0.305 4807 0.441 1221 0.439 1155 0.300 9071 0.136*** 5485 0.134*** 5419 -0.005 96 (0.012) 95 (0.012) 91 (0.022) 88 (0.020) 96 96 96 Firm size: <20 employees 4102 0.191 4499 0.213 1148 0.159 1109 0.170 8601 0.022 5250 -0.032* 5211 -0.021 96 (0.010) 95 (0.012) 91 (0.015) 88 (0.020) 96 96 96 Firm size: [20,200) employees 4102 0.087 4499 0.078 1148 0.075 1109 0.068 8601 -0.009 5250 -0.012 5211 -0.019 96 (0.007) 95 (0.006) 91 (0.015) 88 (0.012) 96 96 96 Firm size: ≥200 employees 4102 0.228 4499 0.176 1148 0.170 1109 0.235 8601 -0.052*** 5250 -0.058*** 5211 0.007 96 (0.012) 95 (0.009) 91 (0.020) 88 (0.021) 96 96 96 Notes: Significance: ***=.01, **=.05, *=.1. Errors are clustered at federal state-year level. Observations are weighted using variable phrf as aweight weights. 34 Figure 7: Regional distribution of the job search methods, sample splitting by sex. 35 9.2 Main variables description This section of the appendix describes the construction of the main variables which are used in the empirical analysis. Current job market status and transitions We use variable pgjobch to define the current job market status of a respondent. It is further augmented by using information on the annual working time: if the respondent reported working more than 52+ hours, s/he is considered to be employed. Based on this information we create the following categories: job-to-job transition, unemploymentto-employment transition, non-participation-to-employment transition, first time participant. To construct a reliable measure for unemployment we additionally used variable plb0021 variable. If in the previous year the respondent was registered unemployed (variable plb0021) and in the current year the status is “Employed (with/without/unclear) job change”, we counted is as a transition from unemployment to employment. First time participants are excluded from the sample. Jobsearchmethodcoding The job search method is the central variable of the analysis (“treatment”), however, it is also the one which considerably restricts the sample size. Respondents undertaking a paid activity were asked if s/he has found a new position in the period from the Jan, 1 previous year till the time of the survey. The job change includes both change of the position within one company and finding a new employer. If this question was answered affirmatively, a respondent was asked about the way which lead to the current position, i.e., the job search method. As SOEP covers a long time period, the ways to find a job have changed several times. For example, in the year 2014 networking, as a job search method, was split by the network types - friend, family or colleagues. Hence, we use a harmonized variable covering 18 possible ways to find a position. These 18 categories were aggregated to the four main job search methods: 1. Formal centralized approach (agency) includes those, who have found a position using the help from an employment office (Arbeitsamt), Job-Center/ARGE/social office (Sozialamt), personnel service agency (Personalserviceagentur) or private job placement. 2. Formal individual search covers the job postings in newspapers and the Internet. 3. Informal individual search includes finding a position via a social network or using social capital of friends, relatives, acquiescences or former colleagues. Thus, it is mostly referred to as networking. 4. All other types of the job search are collected into “other” category. 36 For the analysis in this paper only those respondents who have found their position using formal individual approach (referred to as “individually” or “comparison group”) and with the help of their social capital (“networking” or “treatment”) are included. All variables used for estimation and balance are briefly described in Table 14 below. First column shows a variable’s name in the data/regressions, while the second column provides a short description of the economic meaning. The next column indicates, if a variable was taken directly from the dataset or was constructed by the authors. The forth column indicates, if a variable is a dummy coded as 0-1. Finally, the last column provides notes on construction of the variables. Table 14: Description of variables used age Age, years agesq Age squared, years Yes bula Federal state bula ew Federal State According to statistic Yes East-West Version educ stage Finished stage of education Yes Aggregation of the yerseduc variable. <12 - less than high school; [12,13] - high school lvl; >13 - more than high school educ stage 1 Less than high school Yes Yes Based on educ stage educ stage 2High school Yes Yes Based on educ stage educ stage 3 More than high school Yes Yes Based on educ stage empl change A job-to-job transition Yes Yes See a detailed description firm size A size of the firm # of employees, based on pgbetr variable firm size 1Very small: (0, 20) employees Yes Yes Based on firm size firm size 2 Small: [20, 100) employees Yes Yes Based on firm size firm size 3Middle: [100, 200) employees Yes Yes Based on firm size firm size 4 Large: [200, 2000) employees Yes Yes Based on firm size firm size 5Very large: ≥2000 employees Yes Yes Based on firm size fulltime Working full-time Yes hh size #ofHHmembers hours work year Annual work hours kita less ch The youngest child is in kita age or smaller Yes Yes Based on kidgeb01 - kidgeb15 variables and syear variable. First, the current age of a child is defined. Then for each year the age of the youngest child is saved. Finally, a dummy variable is created, indicating if the child is below 6 years of age in a given year. Variable Name Short Description Constr. A 0-1 dummy variable Notes Continued on next page 37 Table 14: Description of variables used (Continued) log labgro Logarithm of gross montly wages Yes Basedonthelabgrovariable. First, it was winsorized at 1 and 99 percentiles by the type of the occupation. Secondly, a logarithm was taken. long unempl Long-term unemployment, share per county From INKAR mar status Married, living together Yes Yes Based on the d11104 nace full 1d NACE rev 1.1, created, 1 digit Yes Based on pgnace and pgnace2 variables. The first variable is available from 2002 to 2017 and is coded according to the first review of NACE classification. The second variable is available from 2013 to 2019 and based on the rev. 2 of NACE classification. Both variables are aggregated to the major groups and converted to rev 1.1 according to the NACE guidelines. nace full 1d 1 Agriculture/fishing Yes Yes As per NACE rev 1.1 nace full 1d 10 Real estate/business cativities Yes Yes As per NACE rev 1.1 nace full 1d 11 Public administration Yes Yes As per NACE rev 1.1 nace full 1d 12 Education Yes Yes As per NACE rev 1.1 nace full 1d 13 Health and social work Yes Yes As per NACE rev 1.1 nace full 1d 14 Community and social services Yes Yes As per NACE rev 1.1 nace full 1d 15 Private HHs as employers Yes Yes As per NACE rev 1.1 nace full 1d 16 Act. of extraterr. org/bodies Yes Yes As per NACE rev 1.1 nace full 1d 2 Mining Yes Yes As per NACE rev 1.1 nace full 1d 3Manufacturing Yes Yes As per NACE rev 1.1 nace full 1d 4Electr/Gas/Water Supply/Management Yes Yes As per NACE rev 1.1 nace full 1d 5Construction Yes Yes As per NACE rev 1.1 nace full 1d 6 Wholesale/retail trade Yes Yes As per NACE rev 1.1 nace full 1d 7Hotels and restaurants Yes Yes As per NACE rev 1.1 nace full 1d 8 Transport and Communication Yes Yes As per NACE rev 1.1 nace full 1d 9Financial Intermediation Yes Yes As per NACE rev 1.1 nonpart empl A shift from non-participation in labour market to employment Yes Yes See a detailed description pgerljob Occupation matches training pgerljob 1 Yes, matches Yes Yes Based on pgerljob pgerljob 2No, occ. mismatch Yes Yes Based on pgerljob pgerljob 3 In training Yes Yes Based on pgerljob pgerljob 4No training Yes Yes Based on pgerljob pgerwzeit Tenure At a previous employer Variable Name Short Description Constr. A 0-1 dummy variable Notes Continued on next page 38