Migrant Network and Immigrants’ Occupational Mismatch

Abstract This article defines new measures of horizontal and vertical occupational mismatch based on the difference between the skill content of occupations in which individuals have a self-assessed productive advantage, and that of their actual job. It then investigates the impact of network use to find a job on occupational mismatch in the case of immigrants, using original survey data collected among Senegalese immigrants in four host countries. Estimation results show that migrants who obtained their job through the migrant network have a lower probability of negative vertical mismatch. By contrast, network use is not found to significantly affect horizontal mismatch.


Introduction
Occupational mismatch is a particularly acute issue for immigrants in host countries. According to the European Social Survey data covering 22 European countries and 76 immigrants' countries of origin for the 2000-2009 period, 22 per cent of immigrants are overqualified for their job, while this is the case of only 13 per cent of natives (Aleksynska & Tritah, 2013). Whereas immigrants' economic performance in destination countries is partly explained by migrants' self-selection (Borjas, 1987;Chiswick, 1999), occupational mismatch of immigrants is also obviously related to the recognition at destination of the skills that they acquired in their country of origin. Using United States census data, Mattoo, Neagu, and Özden (2008) show that a large part of the occupational underachievement of immigrants in the United States can be explained by the characteristics of their country of origin. This article adopts a complementary approach and explores the occupational performances of immigrants from the same origin in different host countries.
Given the role played by migrant networks in the economic integration of immigrants in host countries, networks very likely affect the occupational performance of immigrants. The impact of migrant networks on job-search and labour market outcomes has led to a sizable literature. In particular, Munshi (2003) shows that the size of the Mexican network in the United States increases the probability for a Mexican immigrant to be employed and get a job in the non-agricultural sector, using rainfall in Mexico to instrument for network size at destination. Amuedo-Dorantes and Mundra (2007) find a positive impact of networks on wages of Mexican immigrants in the United States.
However, to the best of my knowledge, no article has yet specifically addressed the impact of migrant networks on immigrants' occupational mismatch. unobserved characteristics that correlate with their labour market outcomes. For example, if migrants who turn to their network to find a job have on average higher (respectively lower) unobservable skills, such as communication or social skills, they may then be less likely (respectively more likely) to experience negative mismatch, thus resulting in biased estimates of the impact of network use on occupational mismatch.
In order to overcome potential endogeneity problems, I estimate a bivariate probit of network help and occupational mismatch, in which the impact of network help to find a job on mismatch is identified by excluding whether the migrant is part of the Wolof ethnic group from the mismatch equation. This strategy exploits the rich information on migrants' ethnic group collected in the MIDDAS survey and the historical features of Senegalese migration. The Wolof, because of their recent migration history are expected to have smaller networks, here defined as family members or migrants from the same place of origin. I indeed find a negative correlation between being a Wolof and the probability of using the migrant network to find a job.
My identifying assumption is that ethnicity has no direct effect on mismatch. This assumption is further discussed in Section 3. Note already that the exclusion restriction may not hold if ethnicity is correlated with immigrants' unobserved characteristics (for instance attitudes to work) which would also affect occupational mismatch. In order to proxy for these unobserved characteristics, I additionally control for migrants' religious affiliation by including a dummy for Mouride brotherhood.
When the potential endogeneity of network use in the job-search process is not accounted for, I find that migrants who obtained their current job through the migrant network are less likely to suffer negative vertical mismatch in both regions, Europe and Africa, and are more likely to experience horizontal mismatch, though the latter effect is significant in the two African countries only.
When tackling the endogeneity issues by estimating a bivariate probit model on the sample of immigrants in Europe, results are confirmed: obtaining one's job through the migrant network decreases the probability of negative vertical mismatch while it is not significantly correlated with horizontal mismatch.
This article relates to the empirical literature on the impact of migrant networks on immigrants' labour market outcomes in host countries. Although substantial research has been carried out on the impact of migrant networks on employment or wages (Munshi, 2003;Amuedo-Dorantes & Mundra, 2007), no paper has yet addressed the impact of migrant networks on immigrants' occupational mismatch.
In addition, this article makes two contributions to the literature on occupational mismatch. First, it provides the first direct empirical estimation of the causal impact of networks on occupational mismatch, building on the recent job-search theoretical literature. Indeed, the link between networks and occupational mismatch is at the core of the model presented by Bentolila, Michelacci, and Suarez (2010). The basic intuition of their model is that jobs obtained through social contacts may not be perfectly adapted to individuals' skills, given the limited number of occupations in which individuals have contacts who are likely to refer them for a job. Workers with smaller networks may thus choose to lose their productive advantage and accept a job offer through their network in order to reduce unemployment duration. But although their theoretical model is focused on the impact of networks on productive mismatch, they cannot provide a direct test of this effect for lack of appropriate data and indirectly test their predictions by studying the impact of network use on wages. Consistent with their theoretical predictions, they find that jobs found through personal contacts return lower wages than jobs obtained through formal methods and explain this result by a poorer quality of the match.
Using survey data containing information on both individuals' actual occupation and self-assessed productive advantage, I am able to directly test the predictions of the model by Bentolila et al. (2010), by exploring the relationship between network referrals and occupational mismatch, in the particular case of immigrants.
Second, this article contributes to the occupational mismatch literature by defining new concepts of horizontal and vertical occupational mismatch based on the difference between individuals' self-assessed productive advantage and their actual occupation. The mismatch literature indeed almost exclusively focuses on education-based measures of vertical mismatch: mismatch is then defined as over-or undereducation (Chiswick & Miller, 2009;Aleksynska & Tritah, 2013). Even the very few studies that address the issue of horizontal mismatch similarly define mismatch based on education: horizontal mismatch captures the difference between an individual's field of education and her occupation (Robst, 2007;Nordin, Persson, & Rooth, 2010;Nieto, Matano, & Ramos, 2013).
Moreover, to my knowledge, no study has considered simultaneously vertical and horizontal mismatch. The latter, though given much less attention in the literature, may have important economic consequences. Since occupation specific skills are found to increase wages (Shaw, 1984), having a job which does not match one's productive skills is expected to have a negative impact on wages. Very few papers have explored this issue empirically, and only in developed countries. Robst (2007) tests in the United States context the hypothesis that horizontally mismatched workers (based on the comparison of employment and the field of study in college) have lower wages, and finds that mismatched workers earn less than adequately matched ones, once controlling for education level. Nordin et al. (2010) find an even larger income penalty for individuals whose field of education does not match their occupation in Sweden than was found in the United States context. This article more generally relates to the vast literature on the impact of migrant networks on job-search and labour market outcomes. The impact of networks on job-search and labour market outcomes has been much investigated since the seminal research by Granovetter (1995). Networks first matter in the jobsearch process because they are expected to reduce asymmetries of information (Montgomery, 1991). They should thus improve the quality of the match between hiring firms and applying workers, though this view has recently been challenged by the above mentioned paper by Bentolila et al. (2010). Dustmann, Glitz, and Schönberg (2011) for example, who explore both theoretically and empirically the causes of ethnic segregation at the firm level, show that referrals by members of the same minority group reduce uncertainty about the productivity of job market candidates and explain higher wage offers made by the firm. While initial works were focused on developed countries, recent studies have emphasised the role played by networks in developing countries, where they would be used by employers as a screening mechanism (Iversen, Sen, Verschoor, & Dubey, 2009).
The theoretical predictions regarding the role of networks in job-search models are mostly based on a comparison between formal and informal search methods (networks), and depend on assumptions about: the rate of arrival of job offers, which is for example assumed to be lower through the network than through formal search methods in Kugler (2003); the relationship between the strength of the network and the rate of arrival of job offers, which is assumed to be positive in (Goel & Lang, 2009); the possibility for workers to use only one job-search method or compare offers obtained from different channels (as in Goel & Lang, 2009); the relative cost of network search which is assumed to be lower than formal methods search for example in Cahuc and Fontaine (2009). Regarding the different channels for network effects listed above, some specific characteristics of migrant networks can be emphasised. The rate of arrival of job offers through the network may be higher and the relative cost of network search compared to formal search lower for immigrants. Indeed, access to formal search methods may be made more costly by cultural distance and lead immigrants to rely more on informal network-based job searches. Moreover, rejecting a job offer provided by personal contacts may be more costly than rejecting a formal offer, especially for immigrants since they may rely on their networks not only to get a job but also, for example, to find housing or overcome cultural barriers.
The rest of the article is organised as follows. Section 2 describes the MIDDAS survey data and defines the mismatch and network variables. Section 3 presents the empirical strategy and discusses identification issues. Section 4 reports and discusses the results and Section 5 concludes.

Description of the survey
The data used in this article come from four surveys conducted among Senegalese migrants in 2009 and 2010 as part of the MIDDAS project. 2 300 Senegalese were surveyed in each of the top four destination countries of Senegalese migrants, according to the 2012 United Nations Database on international migrants' stocks: France, Italy, Mauritania and Côte d'Ivoire. The sampling design was based on the latest census data in each country to ensure sample representativeness with regard to location, age, and gender. The same fieldwork procedures and questionnaires were used in each country. We collected in particular detailed information on migrants' employment characteristics, wage, education, and their networks and contacts. A more detailed description of the MIDDAS migrant surveys' design is provided in the Online Appendix. This article uses data on labour market outcomes and network use collected on the subsample of around 1000 employed Senegalese migrants.
Regional unemployment data are included as additional controls in some specifications for the subsample of migrants in Europe. Data on unemployment rates at the department level for France (département) come from the French National Statistical Institute (INSEE), and data on unemployment rates at the province level for Italy come from Eurostat. 3

Variables
Occupational mismatch is identified based on the differences between migrants' answers to two distinct open-ended questions: a first question about their profession or trade, and a second question about their actual job. 4 The first question is intended to capture individual comparative productive advantages and its formulation invites migrants to answer according to their self-assessed productive skills. 5 The first step in the construction of the mismatch variables used in this article consisted in ex-post coding interviewees' responses to the two questions about their profession or trade and their actual job using the 2008 International Standard Classification of Occupations (ISCO-08), which classifies all jobs in 436 unit groups at the most disaggregated level. Each unit group is then assigned a skill level by the ISCO-08, as shown in Table A1 in the Online Appendix. The ISCO-08 skill level classification ranges from 1 (elementary occupations) to 4 (highest skill level). 6 Comparing simultaneously the ISCO-08 code and skill level associated with migrants' self-declared comparative advantage (métier) and actual job, I define a variable of vertical (negative) mismatch, which equals one if the migrant's actual job corresponds to a lower skill level than his declared comparative advantage and a variable of horizontal mismatch, which is equal to one if the migrant's actual job does not match her self-declared 'profession' or 'trade' (that is if they have different ISCO-08 codes) but both are ascribed to the same skill level.
Vertical and horizontal mismatch can be illustrated by the two following examples: a migrant who declares being an accountant (skill level 4) and is actually employed as a workman (skill level 1) experiences a negative vertical mismatch. However, being a self-declared tailor and working as a cook is regarded as horizontal mismatch (both are classified at skill level 2). 7 Note that discrepancies between the migrant's self-assessed productive advantage and actual job may capture in part a subjective inadequacy between a migrant's qualification and her occupation. 8 The distribution of skill levels corresponding to migrants' self-declared productive advantage and actual job are represented in Figure A1 in the Online Appendix. Moreover, the frequency tables of the ISCO-08 codes and titles for the 15 most frequent self-declared comparative advantages and actual jobs in our sample are presented in the Online Appendix (Table A2(a) and A2(b)). As appears in these tables, a large majority of interviewees' occupations (either self-declared as being their comparative advantage or corresponding to their actual job) belong to trade and craftsmanship and do not raise any ambiguity as regards their classification (see note seven for the specific case of sales workers).
The proportion of migrants experiencing different types of mismatch, for the separate European and African sub-samples, is shown in Table 1. 9 The share of horizontal mismatch is very similar in both host regions, whereas negative vertical mismatch is much more frequent in Europe (12.7%) than in Africa (2.2%). 10 As for network use, I construct a dummy variable, denoted Network Help or NH, which equals one for migrants who obtained their current job through the migrant network. The migrant network is here defined in a broad sense, including both family members and non-related Senegalese. 11 Table 2 shows the percentage of migrants in Europe and in Africa who obtained their current job with the help of their network. Table 2 presents summary statistics of individual demographic characteristics and network use of employed migrants, by destination (Europe and Africa). The main differences between migrants in Africa and in Europe concern education: the share of migrants with medium to high levels of education is much higher in France and Italy than in African countries, which explains the observed low percentage of negative vertical mismatch in Mauritania and Côte d'Ivoire, shown in Table 1 (2.2%). Indeed, negative vertical mismatch is negatively correlated with education since migrants with a low educational attainment are more likely to declare that they have a comparative productive advantage in a job requiring one of the lowest skill levels, and are therefore unlikely to have an even less skilled job.

Simple probit
As a first step, I explore the potential impact of the migrant network on the probability to experience occupational mismatch by estimating the following equation by a standard probit: Notes: χ 2 /t/Fisher mean or proportion tests in column (3). *p < 0.10, **p < 0.05, ***p < 0.01

Migrant network and immigrant's mismatch 1811
where M Ã i denotes the latent mismatch status of migrant i and is only observed as: where NH i is the network help dummy, and X i is a set of control variables including age, gender, and education. δ j are country fixed-effects. M i is here a generic notation for mismatch but two sets of equations are separately estimated, the dependent variable being alternately a dummy variable for horizontal and negative vertical mismatch, the alternative being the absence of horizontal (respectively, negative) mismatch. However, coefficients estimated by a standard probit are likely to be biased if unobserved variables drive both the probability to find a job through the network and the probability of mismatch.

Bivariate probit
In order to overcome potential endogeneity problems, I jointly estimate the network use and mismatch equations by maximum likelihood. Assuming that the errors in both equations are jointly distributed as bivariate normal, the likelihood has a bivariate probit form, which accounts for the possible endogeneity of network use. 12 With the same notations as above, NH Ã i and M Ã i being latent continuous variables and NH i being an indicator variable for having obtained one's job through the migrant network, which equals one when NH Ã i > 0. Similarly, the indicator variable M i for mismatch, alternately horizontal mismatch and negative vertical mismatch, equals one when M Ã i > 0. ν i and i are individual error terms.

Identification issues
In order to identify the parameters without relying on non-linearities, at least one variable Z i in the network use equation must be excluded from the mismatch equation. I exploit survey information on migrants' ethnicity to construct a dummy variable that equals one for migrants belonging to the Wolof ethnic group. First, the Wolof dummy is expected to be negatively correlated with network use. Indeed, the Wolof are the largest ethnic group in Senegal, with 43 per cent of the population but their migration history is more recent than that of other Senegalese ethnic groups. Senegalese migrations have deep historical roots but the first region of Senegal to participate in massive international migration was the Senegal River valley adjoining Mauritania and Mali, mainly populated by Halpulaar'en and Soninké, as documented by Clark (1994). The Wolof are thus expected to have smaller networks, here defined as family members or friends from the same area of origin, than other Senegalese ethnic groups, in particular in France, which is the historic host country of Senegalese migrants. Interestingly, recent findings by Beaman (2012) using data from refugees resettled in the United States suggest a nonmonotonic impact of network size on labour market outcomes of refugees, depending on the vintage of other network members. Having access to a larger network of recently arrived migrants would deteriorate labour market outcomes through an increase in competition among network members. Such a mechanism reinforces the identification assumptions since the Wolof are both more likely to have access to smaller networks and to suffer from the competition with recently arrived immigrants. Second, the Wolof variable satisfies the exclusion restriction if being a member of the Wolof group has no direct impact on the probability of job-market mismatch. Note that ethnicity could directly affect mismatch in host countries mainly through the following two channels. First, Wolof migrants may differ from migrants from other ethnic groups and have particular unobserved characteristics that correlate with labour-market outcomes. Identification of the causal impact of network use on mismatch relies on the assumption that belonging to the Wolof ethnic group may affect migrants' probability of experiencing occupational mismatch in Europe only through network effects.
Note however that this assumption is less strong than it first seems: it does not imply that ethnicity is uncorrelated with migrants' choice of sector or activity. Belonging to the Wolof ethnic group may indeed be correlated with unobserved characteristics affecting a migrant's attitude to work in general. What is assumed is that being a Wolof has no effect, other than through the migrant network, on the probability of mismatch, that is, on the observed difference between a migrant's self-declared productive comparative advantage (which may well be affected by her ethnic origin) and her actual job.
Moreover, in order to proxy for potential unobserved characteristics of the Wolof that would be correlated with both network use and mismatch, I exploit available information on migrants' religious brotherhood affiliation contained in the MIDDAS data. I include as an additional control variable in both equations of the bivariate probit model above described a dummy variable that equals one for migrants belonging to the Mouride brotherhood. Indeed, socio-anthropological studies document the strength of network links in the widespread Mouride diaspora, and the distinctive work ethic of members of the Mouride brotherhood (see for example Bava 2003).
Second, Wolof migrants may be offered different kind of jobs by employers. This could be the case in particular in Mauritania, since the Wolof account for almost 8 per cent of the Mauritanian population: cultural proximity or a common language could thus explain different labour-market outcomes of the Wolof. The exclusion restriction is less likely to be satisfied for African destinations than for European countries, I thus focus on the subsample of migrants in France and Italy when estimating the bivariate probit model. 13

Probit estimates of horizontal and negative vertical mismatch
Results from standard probit estimations of horizontal and negative vertical mismatch equations are presented in Table 3. Migrants from all four surveyed countries are pooled. These results suggest that network help has a contrasted impact on the different types of mismatch: indeed, finding a job through the migrant network seems to increase the probability of horizontal mismatch but to decrease the probability of negative vertical mismatch. 14 Note that at this stage, we cannot exclude a substitution effect, migrants choosing to accept a network offer corresponding to a horizontal mismatch in order to avoid negative mismatch. Being a man and age at arrival in the host country are positively and significantly correlated with horizontal mismatch. Moreover, horizontal mismatch is significantly higher in Italy than in France. This result is very likely explained by the greater specialisation of Senegalese migrants in Italy in informal trade. The significantly lower probability of negative vertical mismatch in Mauritania and Côte d'Ivoire than in France is a direct consequence of the smaller share of migrants with secondary and tertiary education in Africa. Table 4 presents the same regressions depending on the host region: coefficients and marginal effects of probit regressions for horizontal and negative vertical mismatch on the subsample of migrants in France and Italy are presented in column (1) to (4), and results for horizontal mismatch in African countries in column (5) and (6). Due to the small proportion of negative vertical mismatch in Mauritania and Côte d'Ivoire, the equation for negative vertical mismatch in Africa cannot be estimated. The breakdown of migrants according to their location suggests that network help has a significant positive impact on horizontal mismatch in Africa only.
Alternative specifications (not shown here) were tested, including unemployment rate at a disaggregate level (region or department) as an additional control to account for labour market conditions at the time the migrant first arrived in the host country. This could be done on the French and Italian sub-samples only, due to data availability. I found no significant correlation between the regional unemployment rate and the Migrant network and immigrant's mismatch 1813 probability of experiencing mismatch, and controlling for unemployment does not change the coefficients on network help. This result is consistent with the absence of long-term effects of unemployment at arrival on immigrants' labour market status found in Chiswick, Cohen, and Zach (1997).
It could be argued that host language proficiency should be included in the set of control variables. Indeed, as shown by Dustmann and Van Soest (2002), language proficiency increases immigrants' earnings. However, although information on the language spoken at home is available in the MIDDAS survey data, it cannot be exploited for three reasons: first, it is at best a very imprecise proxy of language proficiency. Indeed, answers are heavily dependent on the migrants' household structure in the country of destination. 15 Second, migrants in the four countries of the survey cannot be compared in that respect. A large majority of migrants in France speak French due to the historical links between France and Senegal and the place of the French language in the formal education system in Senegal. By contrast, very few Senegalese settled in Italy speak Italian. Language proficiency cannot be controlled for since it is highly correlated with the country dummies in the European subsample. In the two African countries, the official language (Arabic in Mauritania, French in Côte d'Ivoire) coexists with several other languages spoken by large shares of the population, making the notion of host language proficiency more complex to measure and unlikely to be captured by the spoken at home language question. Third, as noted by Dustmann and Van Soest (2002), language proficiency is unlikely to be exogenous, and may be correlated with the same unobservable individual characteristics that drive occupational mismatch. Because of these limitations, I stick with the baseline specification that includes migration duration, education dummies and country dummies which are likely to capture most of the language proficiency effect.
As discussed above, the results from the standard probit estimations are likely to be biased. For example, if migrants' unobserved skills (social skills, motivation, and so forth) are correlated with the probability that they get a job through their Senegalese contacts, then, the coefficient on the network help dummy might be biased. This potential endogeneity issue is dealt with by estimating a bivariate probit model.

Bivariate probit estimation results
Bivariate probit results for horizontal and negative vertical mismatch are presented in Tables 5 and 6. Both tables report estimated coefficients and marginal effects for the mismatch and network help equations. The first two columns in each table correspond to the mismatch equation, and the last two columns to the network help equation.
Consistent with the simple probit estimation results, results shown in Table 5 suggest that the impact of network help on horizontal occupational mismatch is not significant for Senegalese immigrants in Europe. Note that being a Wolof is associated with a 12 percentage point decrease in the probability of obtaining a job through the migrant network. This finding is consistent with the identification assumption discussed in Section 3.3, relying on the historical and ethnic characteristics of Senegalese migration. Table 6 presents results for negative mismatch: the impact of network use on the probability of mismatch is found to be large and significant: having obtained one's job through the migrant network is estimated to decrease the probability of negative mismatch by 42 percentage points. Again, as expected, the Wolof dummy is negatively correlated with the use of the migrant network to find a job. Notes: Robust standard errors in parentheses. + p < 0.15, * p < 0.10, ** p < 0.05, *** p < 0.01. The alternative in each column is respectively the absence of horizontal and negative mismatch. (d) for dummy variables. ME are estimated average marginal effects. The reference category for education is no or primary education.
Interestingly, the direct impact of education on negative mismatch is not found to be significantly different from zero. By construction, we expect migrants with secondary or tertiary education to be more likely to experience negative mismatch than migrants with no or primary education. Indeed, they are more likely to declare having a comparative advantage in an occupation associated with a high skill level, which mechanically increases the probability of their actual job being associated with a lower skill level. However, higher educated migrants are also more prone to using formal job search channels (in particular state agencies and competitive exams) rather than relying on their network to find a job. According to the results shown in Table 6, the latter indirect effect dominates. Once controlling for network help, the impact of education on negative mismatch is not significantly different from zero. Unfortunately, the limited size of the sample does not allow me to explore potential heterogeneities in the impact of network help depending on education. 16 Note that the correlation factor between the two error terms in the bivariate probit models is not statistically different from zero for horizontal mismatch and significant at the 12 per cent level for negative vertical mismatch. It is thus not a surprise that the bivariate probit results confirm the results of the estimation of standard probit models for horizontal mismatch, presented in Table 3. For vertical mismatch, the size of the coefficients (and marginal effects) of the network variable is much larger in the bivariate probit specification, suggesting that when omitting to account for the endogeneity of the network variable, the beneficial impact of networks on the quality of the match is underestimated. This latter finding thus indicates that the use of networks in the job-search process might be negatively correlated with unobserved ability. However, the absolute size of the effects should be interpreted cautiously, given the relatively small size of the estimation sample. Both probit and bivariate probit Notes: Standard errors clustered at the region (Italy, Mauritania and Côte d'Ivoire) or department level (France) in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. ME are estimated average marginal effects. The alternative in each column is respectively the absence of horizontal mismatch and network help. The p-value for the correlation factor is obtained after a Wald test. The reference category for education is no or primary education.
results however converge to show that finding a job through the migrant network significantly decreases the probability of negative mismatch. In sum, obtaining a job through the migrant network is found to decrease the probability of a drop in social or occupational status, which is consistent with the strand of the theoretical literature modelling the fact that personal referrals reduce uncertainty about migrants' productivity and increase the quality of the match (Montgomery, 1991). In particular, this finding is consistent with the results of Dustmann et al. (2011) on ethnicity-based networks, who predict that obtaining a job through a referral improves the quality of the match between employees and firms, and find that an increase in the share of workers from their own minority group in the firm is associated with higher wages. The above results on vertical mismatch are also in line with previous findings in the migration literature that emphasise the positive impact of the migrant network on immigrants' labour market outcomes, though not specifically addressing the issue of occupational mismatch. For example, Munshi (2003) finds that migrant networks in the United States decrease Mexican immigrants' probability to be unemployed, and increase their probability to get a nonagricultural job (better paid than an agricultural one). In the same context of Mexican migrants in the United States, Aguilera and Massey (2003) find that network ties positively influence earnings at destination, especially for those who are the most vulnerable (undocumented migrants). In the MIDDAS survey data used in this article, specific information could not be collected on immigrants' legal status because of the sensitivity about this question in the climate of political tension at the time the surveys were conducted in France and Italy. However, the share of undocumented migrants in our European subsample could be estimated from information on the year of arrival Notes: Standard errors clustered at the region (Italy, Mauritania and Côte d'Ivoire) or department level (France) in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. ME are estimated average marginal effects. The alternative in each column is respectively the absence of negative mismatch and network help. The p-value for the correlation factor is obtained after a Wald test. The reference category for education is no or primary education. and type of visa to be between 40 and 60 per cent, which may contribute to explaining the particularly large and positive impact of networks on migrants' labour market insertion in Europe.
On the other hand, as regards horizontal mismatch, my findings differ depending on immigrants' destination. While there is no evidence of an impact of migrant networks on horizontal mismatch in the European sub-sample, I find a positive correlation between network use and horizontal mismatch for the two African countries under study. Although the results should be interpreted cautiously in the African case, since they may be affected by endogeneity biases, they seem to be consistent with the theoretical predictions of the model presented by Bentolila et al. (2010): in African host countries, migrants who obtained their job through their network are on average more likely to get a different occupation from the one in which they have a productive advantage. The data used in the article do not allow further investigation of the different possible underlying explanations, such as a higher cost for immigrants to reject a job offer made by the migrant network or the trade-off between finding a job rapidly and finding a job matching exactly one's skills.
However, the coefficient on the network help variable, though positive, is not significant for Senegalese migrants in the two European countries. Obtaining a job through the migrant network thus has an unambiguous positive impact on the quality of the match for Senegalese immigrants in Europe. Indeed, networks decrease the probability of negative mismatch and do not significantly increase the probability of horizontal mismatch. The different influence of network help on horizontal mismatch in Africa and Europe may partly be due to the greater vulnerability of migrants in Europe, a large share of them being undocumented. For them, the positive role of the migrant network (consistent with Aguilera & Massey, 2003) is likely to far exceed the potential adverse effects put forward by Bentolila et al. (2010), whose model does not account for the peculiarities of the situation of immigrants from a developing country in developed host countries. Such difference between the impact of networks in African and European countries could also be explained by the greater cultural distance experienced by Senegalese immigrants in Europe justifying the beneficial role of network referrals which reduce information asymmetries between employers and workers. A next step in the analysis of immigrants' both horizontal and vertical mismatch would consist in exploring the potentially detrimental impact of mismatch on wages. Descriptive statistics suggest that wages are similar for migrants correctly matched and those experiencing horizontal mismatch. However, the identification of a causal impact of mismatch on wages would imply being able to solve still another endogeneity issue. Indeed, migrants' unobserved characteristics are likely to affect both their probability of mismatch and their wage. Moreover, since information on network use is available only for migrants currently employed, I cannot investigate the initial selection of migrants into employment, which may be facilitated by a larger network. Both issues are thus left for future research.

Conclusion
Using data on labour market outcomes of around 1000 employed Senegalese immigrants in four host countries, this article finds evidence of an impact of network use to find a job on occupational mismatch. Based on the ISCO-08 classification, I define horizontal and negative vertical mismatch depending on whether the occupied position and the self-declared occupation in which workers have a comparative productive advantage are associated with similar or different skill levels.
In order to account for potential endogeneity bias due to omitted variables affecting both migrants' probability to get a job through their network and migrants' status on the labour market, I jointly estimate two equations for mismatch and network use by maximum likelihood. Parameters identification relies on the exclusion of migrants' ethnic characteristics from the mismatch equation.
I find that migrants who obtained their current job through their migrant network are less likely to experience negative vertical mismatch, that is, to have a job requiring a lower skill level than that of the occupation in which they have a comparative productive advantage. On the other hand, the impact of network use to get a job on horizontal mismatch is not found to be statistically significant at conventional levels for immigrants in Europe. All the results presented in the article converge to show a positive impact of network use on the quality of the match in the case of Senegalese immigrants in Europe, and are a new step towards a better understanding of the economic integration of immigrant populations at destination. This article indeed suggests new avenues of research on occupational mismatch not exclusively focused on education, and calls in particular for the inclusion of questions on the productive comparative advantage in migrant surveys.
reported. In the subsequent analysis, if information on self-declared comparative advantage ('métier') is missing and information on current occupation is not, the observation is considered 'matched'. However, all results are robust to the restriction of the sample to observations with non-missing responses for both 'métier' and current occupation, as shown for example in Table A3, in Online Appendix, which presents estimation results of the same model as in Table 4, restricted to the sample with non-missing values for both self-declared comparative advantage and actual job. Note that on this restricted sample (233 migrants in Europe and 254 in Africa), 26 per cent and 23 per cent of migrants in Europe and Africa respectively experience horizontal mismatch, and an additional 24 per cent of migrants in Europe experience negative mismatch. The comparison of these figures to the prevalence of immigrants' labour market mismatch found in the literature is made difficult by the fact that existing studies use education-based definitions of mismatch. For example, according to Aleksynska and Tritah (2013), 13 per cent of immigrants from developing countries in Europe are undereducated and 23 per cent are overeducated. Nieto et al. (2013) find that 35 per cent of immigrants in Europe from non-EU countries are overeducated, and that an additional 46 per cent are horizontally mismatched (based on education). 10. Since the number of observed cases of positive vertical mismatch is very low (around 1% of the whole sample), in the subsequent empirical application they are considered horizontal mismatch. However positive mismatch is more frequent in Europe than in African destinations and may be of concern when regressions are run on the European sample only. To test the robustness of the results to the treatment of positive vertical mismatch, regressions were also run on the European sample after the 16 observations corresponding to positive vertical mismatch were dropped. Regression results (not shown) are then virtually unchanged. 11. The network dummy variable is generated by coding answers to the following multiple choice question relative to interviewees' current occupation: How did you get your job? (in French: Comment avez-vous obtenu votre emploi?). NH is equal to one if the interviewee obtained her current job through a family member, a friend or acquaintance from the same origin village or region or from the same religious community, or through a migrant association. Other response categories included in particular the possibility to have obtained the current job through formal assistance (State agency ANPE), competitive examination, or classified ad. 12. The bivariate probit specification is not rejected by a Murphy's goodness-of-fit score test at the 5 per cent level. 13. Since the bivariate probit model relies on strong distributional assumptions, in an alternative estimation strategy, I estimated a standard two-stage instrumental variable model, using the Wolof dummy as an instrument for network use on the French and Italian samples. The coefficient on the Wolof dummy in the first-stage regression was −0.097, significant at the 5 per cent level, and the F-stat was equal to 6.00. Reassuringly, second stage coefficients on network help in both the horizontal and negative mismatch equations were found to have the same sign as the coefficients in the bivariate probit model, although they were not significant at conventional levels 0.540 (standard error 0.476) for horizontal mismatch, and −0.483 (standard error 0.443) for vertical mismatch, the difference in significance being probably due to the lack of precision of the 2SLS estimator. 14. Similar results (in terms of coefficient sign, magnitude and significance) are found when estimating a multinomial probit with three alternatives (no mismatch, horizontal mismatch and negative mismatch). 15. For example, a migrant living with Senegalese from the same ethnic group (for example family members), say Wolof, will very likely report to speak Wolof at home, even if she is perfectly fluent in French. 16. In the subsample of immigrants in Europe, about 27 per cent obtained their highest diploma in a foreign country (that is, not in Senegal). This percentage rises to 53 per cent for immigrants in Europe with a university degree. In alternative specifications, a dummy equal to one if the highest degree has been obtained elsewhere than in Senegal has been included in the set of explanatory variables. This dummy is not significantly correlated with the network variable nor with the mismatch indicator, either horizontal or vertical, and the main results are unchanged (available upon request).