Neighbourhood Ethnic Composition and Social Participation of Young People in England


 We analyse how neighbourhood ethnic diversity and segregation affect adolescents’ social participation in England. We distinguish between participation in ‘purposeful activities’—such as sports and volunteering—and hanging around with friends. We suggest a novel identification strategy to address the problem of endogeneity of ethnic diversity and segregation. We find that ethnic diversity decreases hanging around, while ethnic segregation increases it. No effects on participation in purposeful activities are found.

Britain has become increasingly concerned about whether the co-existence of different ethnicities may affect the quality of social interaction. Data from the Market & Opinion Research International (MORI) Issues Index show that, since the early 2000s, the quality of the relations among different ethnic groups has been perceived as a major issue faced by Britain, which often scores higher than other salient issues such as the state of the economy or the National Health Service. 1 This article investigates the roots of these concerns for England. It studies how social participation of young people aged 14 or 15 years old is affected by the ethnic composition of the neighbourhood where adolescents live.
We pay particular attention to the role of geography in shaping social participation. Our measures of neighbourhood ethnic composition are ethnic diversity and segregation at the local authority district (henceforth 'district') level. 2 District ethnic diversity does not consider the spatial distribution of ethnicities within the district. This spatial distribution is captured by district ethnic segregation. In our case, district ethnic segregation captures how ethnicities are distributed 2460 the economic journal [august into wards within the district. To define social participation we distinguish 'purposeful activities' (e.g., playing sports, taking part in political activities or community work, and joining youth clubs) from 'hanging around with friends'. 3 We describe the effect of district ethnic composition on the demand for social activities by proposing a simple explanatory model where purposeful activities take place in the district and hanging around with friends takes place in the ward. We test its predictions using the Longitudinal Study of Young People in England (LSYPE). Studying the determinants of social participation is important. Social participation is an objective measure of social capital (Alesina and La Ferrara, 2000;Glaeser et al., 2000;Durlauf and Fafchamps, 2005). 4 Not all forms of social participation are equal. Participation in purposeful activities at a young age predicts many positive outcomes in adolescence and adulthood. Purposeful activities improve cognitive and non-cognitive skills (Broh, 2002;Hansen et al., 2003;Del Boca et al., 2017), educational attainment (Long and Caudill, 1991;Barron et al., 2000;Eide and Ronan, 2001;Lipscomb, 2007;Lechner, 2009;Pfeifer and Cornelißen, 2010;Rees and Sabia, 2010;Stevenson, 2010;Felfe et al., 2016) and labour market outcomes (Long and Caudill, 1991;Barron et al., 2000;Kuhn and Weinberger, 2005;Henderson et al., 2006;Lechner, 2009;Stevenson, 2010;Weinberger, 2014;Lechner and Sari, 2015;Deming, 2017;Lechner and Downward, 2017). Hanging around with friends is often claimed to be associated with undesirable outcomes, such as drunkenness and disorders (DfES, 2006). 5 However, we are not aware of any systematic analysis of hanging around with friends, or other social activities with no specific purpose.
The article fills this gap in the literature by drawing attention to the phenomenon of hanging around. We believe that this is important. In Appendix A we discuss the characteristics and the potential consequences of hanging around. Most young people hang around with friends at least once a week. Young people's hanging around decreases adults' well-being. For young people, hanging around with friends is associated with lower human capital accumulation, higher involvement in risky behaviours and more close friends in adulthood. We use propensity score techniques to claim these associations-particularly those on risky behaviours and number of close friends-are likely to be at least partially causal. For comparison, Appendix A also shows that purposeful activities are associated with-and possibly cause-higher human capital accumulation and more close friends in adulthood.
A growing literature in economics claims that ethnic diversity discourages social participation (Alesina and La Ferrara, 2000;Costa and Kahn, 2003;Charles and Kline, 2006). 6 This literature typically assumes that individuals derive greater utility from forming ethnically homogeneous groups. As in ethnically diverse areas, the probability of making a friend of the same ethnicity is lower than in ethnically homogeneous ones, ethnic diversity results in a decrease of social interaction. Studies on the effect of segregation are limited in number and lead to mixed conclu-2019] ethnic composition and participation 2461 sions. Ethnic segregation has been found to increase the willingness to contribute to local public goods (La Ferrara and Mele, 2007), but also to erode generalised trust (Uslaner, 2011(Uslaner, , 2012. To our knowledge, the effect of ethnic segregation on social participation has never been explicitly studied. 7 The article makes a number of contributions to the existing literature on the effect of neighbourhood ethnic composition on social participation. First, it focuses on young people, as evidence suggests that adolescence is when relational skills are formed (Cunha and Heckman, 2008;Cunha et al., 2010) and social participation in adolescence predicts human capital accumulation, risky behaviours, future sociability and labour market outcomes (see Appendix A). Second, it distinguishes between different forms of social participation, thus accounting for the quality of young people's interaction. Third, it provides the first attempt of modelling the effect of segregation (in addition to ethnic diversity) on social participation, by emphasising the geographical level where hanging around and purposeful activities take place. This is particularly meaningful for young people, who are more restricted in how far they can travel from home, compared with adults. Finally, the article addresses the problem of the potential endogeneity of neighbourhood ethnic diversity and segregation with respect to social participation in a number of ways, including by proposing new instruments based on historical events and legislative acts which shaped the sorting of migrants within England.
We find that district ethnic diversity decreases hanging around, while district ethnic segregation increases it. The effects of district ethnic diversity and segregation on purposeful activities are weak or null. The results are robust to alternative empirical specifications. As hanging around is widely perceived as an undesirable behaviour, we conclude that ethnic diversity does not have detrimental effects on teenagers' social participation. Therefore, policies limiting migration in an attempt to preserve ethnic homogeneity are not desirable. Desegregation policies and policies improving the supply of purposeful activities are likely to be more beneficial.

Data
The main data that we use are from the the Longitudinal Study of Young People in England (LSYPE), a cohort study containing detailed information for approximately 15,000 pupils living in England and born between 1 September 1989 and 31 August 1990. The LSYPE is a twostage probability sample with schools as primary sampling units and students within schools as secondary sampling units. 8 This article is based on the first wave of the data, collected in 2004 when the respondents were 14/15 years old (year 9 at school).
Additional data are added to the LSYPE. Data from the 2001 Population Census (the closest census data to wave one of LSYPE) are used to compute indices of district ethnic diversity and 7 Alesina and La Ferrara (2000) discuss a potential effect of ethnic segregation (on top of ethnic diversity) on social participation, but acknowledge that their analysis 'does not deal directly with segregation'. 8 Maintained schools are stratified by deprivation level (measured as number of pupils receiving free school meal). Deprived schools are over-sampled and so are ethnic minority students. Independent schools are stratified by GCSE results in 2003, boarding status and gender of the students. Pupil Referral Units (PRUs) are a stratum of their own. The selection probability of both independent schools and PRUs is proportional to the number of pupils aged 13 enrolled in the institution. Children educated at home are excluded from the sample and so are children attending very small schools (with fewer than 10 pupils for the maintained sector and fewer than 6 for the independent sector), boarders and children who are in England just for education purposes. Response rate at wave one is 74% (66% full interviews and 9% partial interviews). 2462 the economic journal [august segregation and to construct controls at the district level used in the robustness checks. 9 Districts have been chosen to indicate neighbourhoods as they are responsible for cultural and recreational functions, and thus are the relevant geographical level for looking at social participation. Moreover, with a population ranging from approximately 300,000 to 1,000,000 people (with even large cities like Birmingham or Leeds being formed by a single district), districts are likely to cover the area where most of the social participation takes place. Finally, weather data (monthly millimetres of rainfall and hours of sun) from the UK Met Office are added to account for weather conditions. 10
DEFINITION 1 (F INDEX). The fractionalisation index for district j (F j ) is defined as: where s kj is the share of ethnic group k over the population of district j.
F j is constructed as 1 − the Herfindahl- Hirschman (1964) index and can be interpreted as the probability that two individuals randomly drawn from the population of district j belong to two different ethnic groups. F j ranges from 0 to 1, with higher values indicating high diversity.
District ethnic segregation is measured by the 'Duncan and Duncan segregation index', as in Duncan and Duncan (1955); Cutler and Gleaser (1997); and Cutler et al. (1999). D j can be interpreted as a share of people belonging to one of the ethnic groups of district j that should move to another ward within district j (without being replaced) in order to make ward ethnic diversity equal to district ethnic diversity. D j ranges from 0 to 1, with higher values indicating high segregation.  Figure 1 plots district ethnic segregation against district ethnic diversity for the districts included in our main analysis. Given levels of district ethnic diversity are generally associated with a wide range of levels of district ethnic segregation.

Measuring Social Participation
To measure purposeful activities and hanging around we use two LSYPE questions asking young people about the activities they have carried out when not at school in the four weeks before the interview. The first question lists a set of activities including: sport, playing an instrument, going to the cinema/theatre/concert, or partying. The second question lists activities implying an engagement in the community (e.g., going to youth clubs, doing community work) and two forms of 'hanging around' (near home and in the city centre).
We measure participation in purposeful activities through an indicator (PA) equal to one if the respondent reports carrying out at least one activity among: playing sport, taking part in political activities, going to youth clubs, doing community work. PA is zero otherwise. Two types of activities are excluded from PA: (i) those not necessarily requiring interaction among people, such as playing a musical instrument, going to the cinema/theatre and going to a concert and (ii) those not necessarily 'positive' in the definition of DfES (2006), e.g., partying, dancing, watching HA = 1 if the respondent reports hanging around with friends near home. PA = 1 if the respondent reports carrying out at least one activity among: playing sport, taking part in political activities, going to youth clubs, doing community work. Census and LSYPE data. LSYPE data are weighted. N = 274 (excludes districts with fewer than five respondents). Just districts included in the analysis have been plotted. a football match, going to a nightclub, going to the pub, going to the amusement arcade and playing snooker, as they may be associated with risky behaviours, such as fights, drunkenness or illegal substance use. Our measure of 'hanging around' is 'hanging around near home'. It is measured through an indicator (HA) equal to one of the respondent reports 'hanging around near home', and zero otherwise. 'Hanging around in the city centre' is excluded because it might mean something completely different for young people from different areas of the city. 11 Figure 2 plots district level prevalence of PA and HA against district ethnic diversity and segregation. HA decreases with district ethnic diversity (see top left panel). The share of young people hanging around drops from 60% in the least ethnically diverse districts to less than 40% 11 Hanging around in the city centre is analysed in Section 6. C 2019 Royal Economic Society. Notes: LSYPE wave 1, weighted. The sample is the same used in the empirical analysis discussed in Section 3. Income is total gross household income. Standard errors in parentheses.
in the most diverse ones. HA does not seem to vary with district ethnic segregation (see top right panel). No relationship between PA and either measure of district ethnic composition seems to exist, with the average share of those who participate in PA being just above 60% (see bottom panels). The bivariate relationships in Figure 2 must be interpreted with caution, as they can be partially due to respondents' characteristics differing by districts and also affecting social participation. For example, the effects of household economic deprivation might be mistakenly interpreted as effects of neighbourhood ethnic composition (see Letki, 2008;Fieldhouse and Cutts, 2010;Sturgis et al., 2011). This issue is addressed in Table 1, where social participation is broken down by quintiles of total gross household income. 12 Participation in PA increases with household income. The share of those participating in PA is 73% among respondents in the top quintile of household income and only 59% among those in the bottom quintile. These results are driven by participation in sports. In contrast, participation in HA is pretty constant within quintiles of household income. A full set of descriptive statistics for the sample that we use can be found in Table B1 in Appendix B.

Explanatory Model
Our formalisation of the effect of district ethnic diversity and segregation on HA and PA is based on two assumptions. The first assumption is that HA and PA have different geographical scopes. HA is likely to take place near home, as made explicit by the wording of the question we use to measure it. We refer to the ward where young people live as the space where HA takes place. Appendix A shows that HA generally takes place near home and suggests that this assumption is plausible in our case. PA is likely to take place in an area broader than the ward. We refer to the district where young people live as the space where PA takes place. The reasons for this choice are discussed in Subsection 1.1.

2466
the economic journal [august The second assumption is that young people care about the ethnicity of those participating in HA and PA and have a preference for people of their own ethnic group. This preference for one's own ethnic group, at the core of many explanations of social interaction (e.g., Alesina and La Ferrara, 2000;Charles and Kline, 2006), has been documented for adolescents by Currarini et al. (2009). In Appendix A we use information about the ethnicity of the friends of LSYPE respondents to claim this assumption is plausible in our case.
Our explanatory model focuses on the demand side of social participation. At least in the short run, the supply of social activities is given. The role of supply side is discussed in Section 7. Consider a district, divided into two wards, and assume young people come in two types: Greens (g) and Purples (p). There are two social activities: HA (taking place in the ward) and PA (taking place in the district). The young people decide whether to participate in HA and PA based on an expected utility that depends on their expectations about the ethnicity of the participants in the activity. We assume that young people approximate the ethnicity of the participants in HA(PA) with the ethnicity of the young people in the ward (district). 13 Consider Greens and Purples in ward 1. Their expected utility for HA can be written as: n g1 and n p1 are the numbers of, respectively, Greens and Purples in ward 1. τ HA is a parameter ∈ [0, 1] indicating the level of tolerance for the other ethnicity when young people do HA. If τ HA is 0, young people only derive utility from participants of their own ethnicity; if τ HA is 1, young people's utility does not depend on the ethnicity of the participants. Finally, α HAi is an idiosyncratic parameter, capturing individual preferences and costs related to HA. α HAi ∈ [−ᾱ HAi ,ᾱ HAi ], and α HAi is such that at least one person is always willing to do HA. Defining n g2 = N g − n g1 and n p2 = N p − n p1 , the expected utility of HA for Greens and Purples in ward 2 can be written in terms of the number of Greens and Purples in the district (N g and N p , respectively) and the number of Greens and Purples in ward 1 (n g1 and n p1 , respectively): The expected utility of PA depends on the ethnicity of the young people participating in PA, approximated by the ethnicity of the young people in the district. Therefore, it is the same for young people from the same ethnic group, but living in different wards. The expected utility of PA, for Greens and Purples, can be written:

2019]
ethnic composition and participation 2467 where τ PA is a parameter ∈ [0, 1] indicating the level of tolerance for the other ethnicity when young people do PA. τ PA can be greater, equal or smaller than τ HA . α PAi ∈ [−ᾱ PAi ,ᾱ PAi ] is an idiosyncratic parameter, capturing individual preferences and costs related to PA, and such that at least one person is always willing to do PA.
To guide the interpretation of our empirical results, we study how the incentives of doing HA and PA change when: (i) district ethnic segregation changes with district ethnic diversity constant; (ii) district ethnic diversity changes with district ethnic segregation constant. 14 We distinguish the case where HA and PA are substitutes (i.e., young people are willing and are allowed to switch from activities) from the case where HA and PA are not substitutes. If HA and PA are not substitutes, changes in the utility of HA (PA) can only affect the incentives of doing HA (PA). If HA and PA are substitutes, changes in the utility of HA (PA) may change the relative incentives of doing HA and PA, and thus induce young people to switch between activities.
Let us study the effect of changes in district ethnic segregation. We introduce the perfect integration scenario where: (i) district ethnic segregation is zero: the district and both its wards have the same share of Greens and thus the same ethnic diversity (F d = F w1 = F w2 =F ); (ii) Greens and Purples are equal in number in the district (N g = N p ). (i) and (ii) imply (iii) Purples in ward 1 and Purples in ward 2 are equal in number (n p1 = n p2 ), and Greens in ward 1 and Greens in ward 2 are equal in number (n g1 = n g2 ). Condition (iii) implies (3) and (4) become, respectively: EU p2i (HA) = n p1 + τ HA (N g − n g1 ) N g − n g1 + n p1 + α HAi , To study the effects of changes in district ethnic diversity, we introduce the maximum diversity scenario. In the maximum diversity scenario (i) district ethnic diversity is at the maximum achievable level in our case, i.e., F d = 1 2 (Greens and Purples are equal in numbers in the district); (ii) district ethnic segregation is at a generic positive levelD; (iii) Purples in ward 1 and Purples in ward 2 are equal in number (n p1 = n p2 Proofs of propositions 2.a, 2.b and 2.c are shown in Appendix C. Intuitively, a decrease in ethnic diversity increases both the incentives of doing HA and the incentives of doing PA. If τ PA is 'large enough' compared with τ HA , the incentives of doing HA increases more than the incentives of doing PA. 16 This can happen if, when participating in HA, young people care about the ethnicity of those they hang around with, when participating in PA, they care mainly about the purpose of the activity (helping the community, keeping fit). If HA and PA are substitutes, the majority of the young people may want to switch from the activity whose incentives increase less to the activity whose incentives increase more.
The threshold¯ τ depends on the relative magnitude of τ PA and τ HA . We present some simulations for Greens in ward 1, the largest group. We consider a baseline maximum diversity scenario with 20 Greens and 20 Purples, for different levels of district ethnic segregation (0.1, 0.2, 0.3, 0.4, 0.5). Figure 3 shows the difference in the marginal expected utility (with respect to N g ) of HA and PA for Greens in ward 1. In the top panel, τ = 0. The incentives of doing PA increase more than the incentives of doing HA for any district ethnic segregation and N g . In the middle panel,¯ g1 τ > τ > 0 (τ PA > τ HA , and τ PA and τ HA are similar in magnitude). Whether the incentives of doing HA increase more or less than the incentives of doing PA depends on the level of district ethnic segregation and N g . In the bottom panel, τ >¯ g1 τ > 0 (τ PA > τ HA , and τ PA is substantially larger than τ HA ). The incentives of doing HA increase more than the incentives of doing PA for any district ethnic segregation and N g . 17 15 The new benchmark has been chosen because it is compatible with many levels of segregation. In contrast, a benchmark with zero district ethnic diversity has, by definition, also zero district ethnic segregation. Like in the previous case, when n p1 = n p2 , the maximum achievable level of segregation is 1 2 . 16 In what follows, we refer to this situation as the case where τ PA is 'substantially larger' than τ HA . 17 Note that

Empirical Model
We model HA and PA as linear functions of district ethnic diversity and segregation as follows: F j and D j are, respectively, the fractionalisation and the segregation index for the district j = 1,..., m where person i = 1,...,n lives. Thus, β 1 , β 2 , β 3 and β 4 are the parameters of interest.
Finally, x ij is a vector of controls (individual and family characteristics, plus weather indicators, presented in Subsection 3.2) and ξ 1ij and ξ 2ij are composite errors, decomposed as: ζ 1j and ζ 2j are unobserved effects at the district level, η 1i and η 2i are unobserved effects at the individual level, and 1ij and 2ij are idiosyncratic components. Equations (9) and (10) are first estimated via weighted linear probability models (LPM), with standard errors clustered at the district level. 18 In a second approach (IV), ethnic diversity and segregation are instrumented, and equations (9) and (10) are estimated via weighted two stages least squares (2SLS), general method of moments (GMM), and limited information maximum likelihood (LIML), with errors clustered at the district level. 19

Endogeneity of District Ethnic Diversity and Segregation
Identification of causal effects of district characteristics can be problematic. In our case, there are two types of endogeneity. Endogeneity of type I can arise if people select into districts on the basis of some unobserved individual characteristics η 1i and η 2i correlated with F j and D j (e.g., preferences for local public goods, as in Tiebout, 1956). This makes it difficult to distinguish between the effect of the district as such and the sum of the individual effects of its inhabitants.
which implies¯ g1 τ = 0.4, and thus τ >¯ g1 τ . 18 Weights are sampling weights. We use LPM, as linear specifications are very tractable and permit a direct comparison with the IV results. The LPM was also suitable for our case as all the predicted probabilities estimated for HA through the LPM fell in the [0,1] range, while only one probability for PA fell outside that range. Note that equations (9) and (10) are estimated individually, although participating in PA and HA is likely to be a joint decision, and thus 1ij and 2ij are likely to be correlated. Failing to account for this correlation may lead to a loss of efficiency. We chose not to estimate the equations jointly for comparability with the IV approach. Joint and non-linear estimation of (9) and (10) is presented as a robustness check in Section 5. 19 The 2SLS, GMM and LIML estimators have different strengths and weaknesses. The GMM estimator is more efficient than 2SLS in large samples with heteroscedasticity of the error term caused by clustering. However, GMM can be undesirable in small samples where precise estimates of higher moments are difficult (Baum and Schaffer, 2003). LIML are generally preferred in presence of many (potentially week) instruments and in a small sample, where 2SLS can be seriously biased, while the LIML is median-unbiased.  Endogeneity of type II can arise when unobserved district characteristics are correlated with both district ethnic diversity/segregation and social participation, that is when ζ 1j and ζ 2j are correlated with F j and D j . Endogeneity of type I can arise when people select into districts for reasons correlated with both districts' ethnic diversity/segregation and districts' opportunities of socialisation. This requires, for example, that antisocial people disproportionally select into ethnically diverse districts (Putnam, 2007). Endogeneity of type I would also occur if poorer and lower educated people can only afford districts with worse provision of local public goods (e.g., sport facilities, parks). 20 If those districts are also the most diverse/segregated, the correlation between district ethnic diversity/segregation and social participation could be spurious.
Endogeneity of type II can arise when the central government disproportionally invests in purposeful activities in ethnically diverse or segregated districts to engender community cohesiveness. 21 This problem is unlikely to exist in our case, as recreational activities are provided by districts. Indeed, for the case of sport, evidence has claimed that investment in sport undertaken by the central government in the years we consider 'was not primarily aimed at getting the local community involved in sport' (Gratton et al., 2005).
Our empirical strategy limits both types of endogeneity. Endogeneity of type I is limited as young people are less mobile than adults and are not directly involved in parental location decisions. Therefore, households' residential choices are not directly correlated with children's taste for social participation and depend mainly on parental characteristics we can control for. This reduces potential correlations between η 1i and η 2i and the indices of ethnic diversity and segregation. Endogeneity of type II is limited as district ethnic diversity and segregation are measured using 2001 census data (i.e., three years before 2004, when our data on social participation were collected). This reduces potential correlations between F j and D j and the error term due to districts-specific shocks that occurred in 2004.
To address the problem of endogeneity of type I, we include in x ij a set of proxies for preferences and constraints affecting social participation, potentially correlated with F j and D j through endogenous residential choices. At the individual level we control for gender and ethnicity: two important predictors of preferences for and constraints to social participation (Antunes and Gaitz, 1975;Platt, 2009). We also control for whether the respondent was born in 1989 and whether English is not her mother-tongue language. The former captures differences in age and anything else affecting people born in different calendar years, the latter captures ability and willingness to interact with natives. 22 At the household level we control for mother's characteristics (education, employment status and age) and whether the main parent is the mother. 23 These variables capture heterogeneity in parental decisions about young people's leisure time allocation. We also control 20 Appendix A shows parks and green areas are where young people hang around the most. 21 The economic literature generally does not see differences in investment in local public goods by the local government as an endogeneity problem. It claims ethnic diversity and segregation determine both social capital and provision of local public goods (Alesina et al., 1999;La Ferrara and Mele, 2007). In our case, this implies β 1 , β 2 , β 3 and β 4 pick up the total effect of district ethnic diversity and segregation on social participation, including supply side effects through investment in local public goods. For a discussion of the supply side effects, see Section 7. 22 Most LSYPE respondents were born in 1989 or 1990. Nineteen people were born before or after these dates. They are excluded from the sample. 23 Mothers' characteristics are captured through a set of dummy variables. Education variables are: no education, GCSE and below, above GCSE, degree, missing information on education. Employment status variables are: mother working full-time, part-time, not working, missing information on employment status. Age variables are: below 35, 35/50, above 50, missing age. In LSYPE the main parent is the parent most involved in the young person's education, who provides household level information. Fathers' characteristics are strongly correlated with mothers' and their measurement is poor due to high non-response. Therefore, they are not included in the analysis.

the economic journal
[august for household income (in quintiles) to capture availability of resources. Finally, we control for the millimetres of rain and the hours of sun in the month of interview. Our controls may not be enough to avoid any potential source of endogeneity, especially endogeneity of type II. Type IV tackles this.

The instruments
We propose two new sets of instruments motivated by historical research. The first set uses measures of proximity to the ports listed in the Aliens Act: a 1905 piece of legislation restricting the number of ports through which individuals could legally enter Britain (see Pellew, 1989). The second set of instruments is based on measures of proximity to the so-called mill towns. 24 At the end of the Second World War, mill-town jobs were becoming unattractive for native people (Simpson, 2004). At the same time, constraints on migration were loosened, creating an exogenous supply of workers for jobs in the textile industry (Kalra, 2000).These new workers, who worked night hours and weekends and spoke very little English, never mixed with the English majority (Cantle, 2001;Simpson, 2004;Swanton, 2010). These two historical events had different effects on district ethnic composition. Migrants located by the ports were likely to be seamen, inclined to temporary migration and used to travelling and interacting with other ethnicities (Tabili, 1994). Migrants to mill towns were a fairly homogeneous group, spoke poor English, and had relative long-term perspectives of staying in England. This led to high diversity and low segregation in ports and high segregation in mill towns. Subsections 3.3.2 and 3.3.3 discuss instruments' exogeneity and relevance.

Exogeneity
Distance from the main ports of entry has been used to instrument cultural diversity in the USA (Ottaviano and Peri, 2006), following the argument that a port's location is exogenously determined by land morphology. Our strategy rests on a similar assumption and has the further advantage that ports are chosen based on a historical legislative act. This avoids arbitrary ports selection by the researcher and simultaneous determination of port status (determined in 1905) and social participation (determined in 2004). To be exogenous, Aliens Act port status should be uncorrelated with district level unobserved effects potentially correlated with social participation. In other words, cities chosen as ports of entry in 1905 should not be substantially different from other areas in the provision of social activities, or, if such difference existed, it should not persist until 2004. This is difficult to test, as, to the authors' knowledge, there are no data on social activities for the early twentieth century.
To proxy for the provision of social activity we use data on population growth from the Urban Population database 1801-1911 (Bennett, 2012). 25 We want to explore if before the introduction of the Alien Act future ports of entry were already growing faster (in terms of population, and thus development) than other areas. The idea is that high development could be associated with 2019] ethnic composition and participation 2473 high provision of social activities. Figure B1 in Appendix B plots the logarithm of population against time. It shows that, although the Aliens Act ports of entry were on average bigger than other towns or cities, the population growth rate in the nineteenth century (the slope of the line) for the two groups is almost identical. This suggests ports listed in the Alien Act were not very different from other areas of Britain in terms of population growth, and possibly also in terms of unobservables. 26 Proximity to mill towns could not be considered an exogenous instrument if the area where the first mill towns were established was more advanced than other areas in terms of a number of aspects potentially endogenous to social participation, such as technical or scientific knowledge. This is unlikely for three reasons. First, the spin-off of the industrial revolution was the presence, in particular areas across Britain, of a strong watch industry (Allen, 2009) that had only a mediated effect on the British industrial development. The watch industry increased the availability of cheap, good quality gears (and skilled workers capable of assembling them), which facilitated the introduction of crucial 'macro inventions' (Mokyr, 1990;Allen, 2009). 27 Second, the key inventors of the cotton revolution were generally artisans, rather than leading scientists. 28 Third, similar technologies to those incorporated in the 'macro inventions' were already in use in other areas of Britain (Allen, 2009). 29

Relevance
Relevant instruments must predict district ethnic diversity (segregation) in 2004. This condition is satisfied if: (i) the Aliens Act affected the ethnic composition of the areas by the ports (the decline of the textile industry and the contemporary loosening of migration constraints affected the ethnic composition of the areas by the mill towns); (ii) the ethnic composition of Britain resulting from (i) persisted until 2004.
We claim that condition (i) is satisfied for the case of ports of entry, as the Aliens Act was successfully enforced and thus changed the ethnic composition of the areas by the ports (Pellew, 1989;Tabili, 1994;Little, 2013). Condition (i) is satisfied for the case of mill towns, as the spread of the cotton industry outside its original location was promptly inhibited by licensing policies requiring would-be cotton spinning firms to bear the burden of assembling water frames. This favoured producers located in proximity to the watch industry (Allen, 2009). Condition (ii) is satisfied as the migration legislation introduced in the late twentieth century hindered temporary migration and forced new migrants to move where pioneers had already settled down (Holmes, 2001). 30 This created a strong correlation between location decisions of new and old migrants, 26 Note that differences in population growth between Aliens Act ports and other cities would not invalidate our identification strategy, however, similarities in population growth could suggest similarity in other aspects, including development, and thus provision of social activities. 27 The 'Spinning Jenny' (1764), the water frame (1769), the carding machine (1775) and the mule (1779). 28 Other areas in Britain had stronger links with the scientific community, like London, home of the the Royal Society, Birmingham, hosting the 'lunar society', and Edinburgh, with a world leading school of chemistry. Allen (2009) shows that, among all the key sectors of the British industrial revolution, the textile industry was the one with the weakest link with the scientific community. 29 Spinning machines have been documented in the early 1700s in Norwich, spinning and carding machine in Birmingham between late 1730 and late 1750, and carding machines in 1748 in Herefordshire (Allen, 2009). 30 For example, in 1962, the 'Commonwealth Immigrants Act' introduced an entry system for Commonwealth workers based on vouchers. This system was followed in 1968 by a quota system and by further restrictions based on evidence of partiality in 1971.  Notes: Tests and are computed using the ivreg2 stata command by Baum et al. (2002).
with the effect that England's ethnic composition has remained stable throughout the twentieth century (Pellew, 1989;Holmes, 2001). 31 To create our instruments, we compute the inverse of the geodesic distance between each district in England and the closest Aliens Act port/mill town. 32 These measures, together with the variables in x ij , are used to predict district ethnic diversity and segregation. Results are shown in Table B2 in Appendix B. Proximity to ports of entry is positively (negatively) and significantly correlated with the diversity (segregation) index. Proximity to mill towns is positively and significantly correlated with the segregation index only.
For our type IV analysis we allow for a number of non-linearities between proximity to ports/mill towns and district ethnic composition, adding higher orders of the measures of proximity up to cubic. Further, to capture the effect of ports/mills other than the closest, we compute the number of ports/mills within a radius of 20, 50 and 100 km, plus the interactions with the measures of proximity (without higher terms). We call this type IV approach 'IV1'. 33 Tests of instruments' relevance are presented in Table 2 (top two panels). They all suggest that the instruments are relevant. The first panel presents underidentification tests, leading to the conclusion that the model is not underidentified. 34 The second panel presents the Kleibergen-Paap (KP) Kleibergen and Paap (2006); Kleibergen and Schaffer (2007) and the Sanderson-Windmeijer (SW) (Sanderson and Windmeijer, 2016) F-tests on the instruments in the first stages. 35 All F-31 Holmes (2001) documents that the main ethnic groups living in Britain at the beginning of the twentieth century were the main ethnic groups we observe in Table A10 (African, Caribbean, Indian, Bangladeshi, Pakistani and Chinese, with the Chinese community being the smallest in size). 32 The shortest path between two points a, b of coordinates (lat(a), long(a)) and (lat(b), long(b)) on the earth's surface can be computed using the great circle distance, corresponding to the length of a straight line in an Euclidean space. The great circle distance is obtained by multiplying the central angle between the two points (in radians) by the radius of the sphere, i.e., dist ab = arccos(sin(lat(a)) * sin(lat(b)) + cos(lat(a)) * cos(lat(b)) * cos(| (long(b) − long(a)) |)) * 6371, where 6371 is the radius of the earth in kilometres. 33 Slightly different IV approaches (IV2, IV3 and IV4) are presented as robustness checks. 34 The Kleibergen-Paap (KP) underidentification test tests if the matrix of the reduced form coefficients is full rank; that is if the correlations between the instruments and the endogenous variables are different from zero (null hypothesis: the model is underidentified). The Sanderson-Windmeijer (SW) underidentification tests (Sanderson and Windmeijer, 2016) test underidentification regressor by regressor. 35 The SW F-test is considered the most appropriate in the case of multiple endogenous variables. We also computed the Angrist and Pischke (AP) F-statistics (Angrist and Pischke, 2009). With two endogenous variables they converge to the SW statistics, and thus results are not reported. In addition, the comparison between partial R 2 of excluded 2019] ethnic composition and participation 2475 Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English notmother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Weighted data.
tests are greater than 10, providing additional evidence in favour of instruments' relevance.
Instruments' exogeneity is difficult to test. In case of over-identified models, most scholars refer to the Sargan-Hansen test to derive an indication of instruments' exogeneity. Sargan-Hansen tests for both PA and HA (see bottom panel of Table 2) fail to reject the null of orthogonality of the instruments. Table 3 shows the coefficients for district ethnic diversity and segregation estimated via LPM. Four specifications are presented: with district ethnic diversity only (specification 1), with both district ethnic diversity and segregation (specification 2), with individual controls (specification 3), with individual, household and weather controls (specification 4). Specification 4 is our main specification. 36 The estimated coefficient for district ethnic diversity on HA is negative and strongly statistically significant in all specifications. The estimated coefficient for district ethnic diversity on PA is negative in specifications (1) and (2), but goes towards zero when controls are added in specifications (3) and (4). This lack of effect of district ethnic diversity on PA is compatible with our explanatory model in a case where τ PA is substantially larger than τ HA . If HA is motivated by a pure desire of socialisation, while PA is also motivated by other goals (keeping fit, helping the community), young people may care more (less) about other people's ethnicity when doing HA (PA). The estimated coefficient for district ethnic segregation on HA is positive and statistically significant in all specifications. The estimated coefficient for district ethnic segregation on PA is negative but statistically insignificant in all specifications. These results are compatible with our explanatory model.

Results
To have an idea of the magnitude of the effects, consider specification (4). Going from the median (0.094) to the highest observed ethnic diversity (0.676) leads to a decrease of around 11 percentage points (p.p.) in the probability of participating in HA. This negative effect is instruments and Shea's R 2 suggests that both instruments are relevant (R 2 (diversity) = 0.371, Shea's R 2 (diversity) = 0.432, R 2 (segregation) = 0.248, Shea's R 2 (segregation) = 0.289). 36 Estimated coefficients for the control variables are shown in Table B3. They generally suggest young people from richer and more educated households are more likely to choose PA, while young people from more deprived background, but also young people with working mothers, are more likely to choose HA.  Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. IV results are computed using the ivreg2 stata command by Baum et al. (2002). Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Instruments: proximity to the closest port of entry and its higher terms (up to the third), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest port and the number of ports within 20, 50, 100 km. Weighted data.
comparable to the one associated with the dummy variable indicating respondents whose mother tongue is not English (see Table B3 in Appendix B). Similarly, going from the lowest observed ethnic diversity (0.02) to the median leads to a decrease in HA of just above 1 p.p. Finally, going from the lowest observed ethnic segregation (0.058) to the median (0.160) leads to an increase of around 2 p.p. in the probability of HA, while going from the median segregation to the highest observed value (0.620) increases this probability by 9 p.p. This latter effect is larger than the effect of having a mother with no education, as opposed to having a mother with a university degree (see again Table B3). Table 4 shows the estimated coefficients for ethnic diversity and segregation on PA and HA when endogeneity is accounted for. They are similar to those obtained via LPM. 37 The results are robust across estimation methods (2SLS, GMM, LIML). The 2SLS and the LIML estimates are very similar, suggesting our instruments are not weak (Angrist and Pischke, 2009). The estimated type IV coefficients for district ethnic diversity and segregation are generally statistically insignificant for PA and strongly statistically significant for HA. For HA, the coefficients for district ethnic diversity become more negative than the one estimated via LPM, with increases in absolute value of up to 23%. The coefficients for district ethnic segregation become more positive and more than double in magnitude. For example, 2SLS and LIML suggest that passing from the median to the highest observed district ethnic diversity (segregation) leads to a 13 p.p. decrease (23-24 p.p. increase) in HA.
The lack of correlation between PA and district ethnic diversity is coherent with our explanatory model when τ PA is substantially larger than τ HA . Moreover, it is also in line with the critique that part of the negative effects of neighbourhood ethnic diversity on social participation is due to omitted neighbourhood deprivation (see Letki, 2008). This critique is valid in our case if: (i) ethnically diverse districts are the most deprived; and (ii) district deprivation is associated with poor access to PA. Indeed, in the specifications without individual, household and weather controls (Table 3, specifications (1) and (2)), ethnic diversity is found to be negatively correlated with PA, but such correlation disappears when the endogeneity problem is addressed first by adding controls and then through the type IV approach.
For HA, a similar argument explains why controlling for endogeneity makes the coefficient of ethnic diversity even more negative. Suppose that district ethnic diversity leads to less HA.

2019]
ethnic composition and participation 2477 Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Results are computed using the cmp stata command by Roodman (2011). Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Instruments: proximity to the closest port of entry and its higher terms (up to the third), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest port and the number of ports within 20, 50, 100 km. Weighted data.
Suppose also that ethnic diversity is positively correlated with deprivation, and deprivation leads to more hanging around. Thus, deprivation implies young people in more ethnically diverse districts do more HA. This positive bias would mask the 'real' negative effect of district ethnic diversity on HA. Thus, the type IV coefficient of district ethnic diversity on HA is likely to be more negative than the LPM one.
There is no literature on the effects of ethnic segregation on the demand of social activities. On the supply side, La Ferrara and Mele (2007) suggest that district ethnic segregation leads to higher provision of local public goods. If local public goods include parks, security, sport facilities and clubs, this may facilitate participation in both PA and HA. If less social families self-select into more segregated districts, this causal effect may show up only when endogeneity is controlled for. Thus, the type IV coefficient of district ethnic segregation on HA and PA is likely to be more positive than the LPM one.

Robustness Checks
We present a number of robustness checks. First, we test the robustness of the results to the functional form used in the estimation, by estimating the model in (9) and (10) non-linearly. We first estimate the model by probit and then we use a conditional mixed process (CMP), together with IV1 instruments, to account for endogeneity of district ethnic composition. 38 We estimate three different specifications of the CMP. Each specification consists in four jointly estimated equations. Two equations, equivalent to the first stages in the fully linear case, estimate district ethnic diversity and segregation. Two equations, estimated by probit, estimate the probability of taking part in PA and HA. The exogenous variables included in the first stages are: CMP1, IV1 instruments and variables x ij in (9) and (10); CMP2, IV1 instruments without variables x ij in (9) and (10); CMP3, IV1 instruments on ports of entry to predict ethnic diversity and IV1 instruments on mill towns to predict ethnic segregation.
Marginal effects at the mean (in Table 5) are consistent across specifications and similar to those obtained through linear models. The probit estimation suggests district ethnic diversity has a small (marginally significant) negative effect on PA, which disappears once endogeneity is accounted for. The effect of district ethnic segregation on PA is not significant when estimated the economic journal [august   Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. IV results are computed using the ivreg2 stata command by Baum et al. (2002). Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. District level controls: population of the district (divided by 10,000), number of vehicles in household, percentage of unemployed people (16-74 years old), percentage of people without any educational qualification (16-74 years old), average household size, percentage of people in good health, under-18 conception rate (multiplied by 1,000), number of road casualties (divided by population of the district), income deprivation affecting children index (IDACI). Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Instruments IV1: proximity to the closest port of entry and its higher terms (up to the third), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest port and the number of ports within 20, 50, 100 km. Weighted data.
via probit and becomes slightly significant and positive in some specifications accounting for endogeneity of neighbourhood ethnic composition. Results for HA confirm a strong negative effect of district ethnic diversity and a strong positive effect of district ethnic segregation. Second, we test the robustness of the results to small changes in the construction of the instruments. We present three different specifications of the instruments (IV2, IV3 and IV4) in addition to the IV1 specification used in the body of the article. The functional form and the set of controls used remain the same as those used in IV1. IV2 instruments district ethnic diversity and segregation with measures of proximity to both the first and second closest port/mill town, their squares, the number of towns within 20, 50 and 100 km, and the interactions between these variables and the proximity from the first and second port/mill town. IV3 instruments district ethnic diversity and segregation with measures of proximity to all the ports and mill towns. IV4 instruments ethnic diversity and segregation with a series of binary variables capturing the number of ports (mill towns) within a 20, 50 or 100 km radius. Table B6 shows the tests on the instruments are fully satisfactory for all strategies. Second stages, shown in Table B7, are in line with those from our main specification. In conclusion, these robustness checks show the results are robust to modest changes in how instruments are constructed.   Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. IV results are computed using the ivreg2 stata command by Baum et al. (2002). Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. District level controls: population of the district (divided by 10000), number of vehicles in household, percentage of unemployed people (16-74 years old), percentage of people without any educational qualification (16-74 years old), average household size, percentage of people in good health, under-18 conception rate (multiplied by 1,000), number of road casualties (divided by population of the district), income deprivation affecting children index (IDACI). Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Instruments IV1: proximity to the closest port of entry and its higher terms (up to the third), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest port and the number of ports within 20, 50, 100 km. Weighted data.
Third, we test the robustness of our results to the inclusion of additional district level characteristics that are likely to affect HA and PA. Such characteristics are: population of the district (divided by 10,000), average number of vehicles in households, unemployment rate (for people 16-74 years old), percentage of people with no educational qualification (16-74 years old), average household size, percentage of people in good health, under 18 conception rate (multiplied by 1,000), number of all road casualties (divided by population of the district), income deprivation affecting children index (IDACI). 39 We estimate this richer specification (RS) first via linear probability models and then by instrumenting district ethnic diversity and segregation with IV1 instruments. In a further sets of specifications, we also include county fixed effects. 40 RS rests on different assumptions than the main specification of the article. Controlling for a larger number of district level characteristics 39 Data on district population, average number of cars, unemployment, health, education, and household size are from the 2001 Population Census data, while data on road casualties and under 18 conception rate are from the 'neighbourhood statistics' collected by the ONS. IDACI is computed at the Super Output Area and can be obtained with the LSYPE. Data are aggregated by district using sample weights. 40 County is the administrative level higher than the district. We included one dummy for each county in England and a dummy indicating 'unitary authorities': a special type of districts (generally bigger cities) which do not belong to any county. the economic journal [august Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. IV results are computed using the ivreg2h stata command by Baum and Schaffer (2012). Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. District level controls: population of the district (divided by 10,000), number of vehicles in household, percentage of unemployed people (16-74 years old), percentage of people without any educational qualification (16-74 years old), average household size, percentage of people in good health, under-18 conception rate (multiplied by 1,000), number of road casualties (divided by population of the district), income deprivation affecting children index (IDACI). Weighted data.
may reduce the district level endogeneity left in the error term. However, while district ethnic diversity and segregation are allowed to be endogenous (and are instrumented), the other district level characteristics are assumed exogenous. Instrument diagnostics for RS models with IV1 are shown in Tables B8 and B9. The instruments perform slightly worse than in our main specification. However, the tests are overall satisfactory. Results from the RS specifications (both LPM and IV) are shown in Tables 6 and 7. They confirm there is no effect of district ethnic composition on PA and there is a negative (positive) effect of district ethnic diversity (segregation) on HA. These results are in line with those from our main specification, in spite of the differences in the identifying assumptions.
Fourth, we test the robustness of our results to the use of a completely different type IV strategy. We use the instrumental variable approach introduced by Lewbel (2012) and we apply it both to the specification without district level controls and to the richer specification with district level controls, with and without county fixed effects. Lewbel's approach derives identification from the standard assumption of exogeneity of the controls and the additional assumption of heteroscedasticity of the error term. It can be used when valid exclusion restrictions are not available. Results are shown in Table 8: they are in line with those obtained with the other identification strategies. 41 The results in this section and in Section 4 come from models relying on a very diversified set of identification assumptions. 42 In spite of these differences, they all tell the same story. All models suggest that district ethnic diversity (segregation) has a negative (positive) impact on HA. These effects become even larger when endogeneity is accounted for, although most robustness checks suggest the impact of district ethnic segregation on HA is smaller than the one implied by the type IV results in Table 4. Where endogeneity is controlled for, we are generally unable 41 Tests of relevance and exogeneity of the instruments as well as tests on the heteroscedasticity of the error term are shown in Tables B10 and B11. 42 For another paper comparing results from a geographically inspired IV strategy and result from Lewbel's IV strategy, see Emran and Hou (2013). to find negative effects of district ethnic diversity on PA. This is coherent with our explanatory model where τ PA is substantially larger than τ HA . We also find some-rather limited-evidence that there may be a positive effect of district ethnic segregation on PA. In Section 7 we provide evidence suggesting that this result is likely to be driven by supply side effects.

Demand Side: More on Mechanisms
Our explanatory model focuses on the demand side of social participation. It emphasises the role of geography: HA takes place in the ward and PA takes place in the district. Moreover, the model acknowledges that HA and PA may be different in nature. In particular, they may differ in the importance young people attach to the ethnicity of those participating in the activity (captured by the tolerance parameters τ HA and τ PA ). In this section we argue that differences in the tolerance parameters τ HA and τ PA capture a meaningful difference in the nature of PA and HA, and that geography is important in explaining our results.
In Section 4 we claim that the relative magnitude of τ HA and τ PA may explain why district ethnic diversity affects participation in HA more than participation in PA. Moreover, we claim that geography explains why district ethnic segregation affects participation in HA more than participation in PA. Is it possible that the different effect of district ethnic segregation on HA and PA is not driven by geography and, instead, is driven by the nature of PA and HA, in a way that is not captured by the tolerance parameters τ HA and τ PA ? In what follows, we compare two activities similar in nature to claim that it is unlikely that our results on the effect of district ethnic segregation are driven by a difference in the nature of HA and PA not captured by our model, and that geography is likely to play an important role.
We consider a third activity: 'hanging around near the city centre' (HC). HC is important for a number of reasons. First, it is popular among young people: around 30% of our sample reports having done HC in the four weeks before the interview. Second, HC is similar in nature to HA, as it is likely to be motivated by the desire to spend time with friends with no specific goal. Third, due to this similarity, HC is likely to display a high degree of substitutability with HA. Finally, like PA, HC does not necessarily take place in the ward. We estimate the effect of district ethnic diversity and segregation on HC. We are particularly interested in the coefficient for district ethnic segregation. If the results for district ethnic segregation are only driven by the nature of the activity, the effects of district ethnic segregation on HC should resemble the one found for HA. If the geographical scope of the activity also matters, the effect should get closer to the one found for PA.
Results (in Table 9) show district ethnic segregation has a strongly significant positive effect on HA and a negative effect on HC (significant at the 10% level). As previously discussed, district ethnic segregation increases the incentives to participate in HA (which takes place in the ward), but not the incentives to participate in activities such as PA or HC, which may take place elsewhere in the district. If HA and HC display high substitutability, an increase in HA can lead to a decrease in HC. In segregated districts, where wards are on average more homogeneous than the district, people may prefer to hang around near home rather than in the city centre. This suggests that the nature of the activity is not the only driver of the results, and that geography plays a role in explaining the impact of district ethnic composition on social participation. 43 43 The estimated coefficients of interests obtained instrumenting district ethnic diversity and segregation using IV1 are shown in Table B12. They confirm a statistically significant negative effect of district ethnic segregation on HC C 2019 Royal Economic Society. the economic journal [august Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Weighted data.
To investigate the role of geography further, we estimate an augmented model by adding to our main LPM specification the fractionalisation index at the ward level. The coefficient for ward ethnic diversity should be negative for HA (HA is discouraged by diverse wards). Ward ethnic diversity should be irrelevant for PA: its coefficient should be close to zero. 44 This augmented model also helps decompose the effect of district ethnic segregation. The coefficient for ward ethnic diversity captures the effect of district ethnic segregation due to ward ethnic diversity. Due to the inclusion of ward ethnic diversity, the coefficient for district ethnic segregation in the augmented model measures the effect of district ethnic segregation due to the set of available ethnic mix of wards within the district (see La Ferrara and Mele, 2007).
Results are presented in the top panel of Table 10 (last two columns). HA is negatively correlated with ward ethnic diversity, but not with district ethnic diversity. This suggests HA is likely to take place in the ward, and thus ward ethnic diversity is what primarily affects the choice of engaging in HA. The inclusion of ward ethnic diversity leads to a reduction in the coefficient for district ethnic segregation if compared with the one in the second column. This is what we expect, as the coefficient for district ethnic segregation in the second column captures both the effect of ward ethnic diversity and the effect of the available ethnic mix of wards within the district. The estimated coefficient for district ethnic segregation in the augmented model in Table 10 remains positive and statistically significant at the 1% level. In sum, results in Table 10 suggest geography plays an important role in explaining participation in HA and PA.
Ward ethnic diversity may be endogenous to social participation due to endogenous sorting of households into wards. To address this potential problem, we restrict the sample to respondents living in social housing (see Algan et al., 2016;Bonomi Bezzo, 2017). 45 In 2004 social houses were allocated by districts through a waiting list system, and households had very limited and a negative-but in most cases not statistically significant-effect of district ethnic diversity. Tests on instruments relevance, in table B13, show the instruments are valid. Note that the substitutability between HA and HC may also explain the coefficient for district ethnic diversity on HC. According to our explanatory model, social activities (including HC) decrease when district ethnic diversity increases. HC and HA are similar in nature, and thus it is also likely that τ HC ≈ τ HA . However, for some τ HC > τ HA , an increase in ethnic diversity decreases the incentives of doing HA more than the incentives of doing HC. For this reason, young people may switch from HA to HC. This may mitigate the negative effect of district ethnic diversity on HC. 44 Reality may partially depart from our simplified framework. For example, when PA are provided in multiple wards within the districts, young people may choose the activity taking place closer to their house. Therefore, the coefficient for ward ethnic diversity may become slightly negative also in the case of PA. 45 We selected respondents who rented either from the council or from a housing association.

2019]
ethnic composition and participation 2483 Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Weighted data.
discretionality in the choice of where to live. 46 This strongly limits the scope of endogenous sorting into wards. Results for the 'social housing' subsample are presented in the bottom panel of Table 10 (last two columns) and are in line with those for the full estimation sample. This suggests that the results are not driven by endogeneity. 47 46 The allocation of social housing was regulated by the Housing Act 1996. Houses were allocated on a first-come-firstserved basis. Priority could be given to subjects who were homeless, lived in unhealthy or overcrowded accommodations, were ill or disabled. Preferences for location were taken into account only when a failure to relocate would have caused hardship for the subject or for others. Houses could also be allocated to avoid the concentration of residents with similar socio-economic characteristics, such that the greatest possible social mix could be achieved. With the introduction of Choice Based Letting (CBL) schemes, the Homelessness Act 2002 made it possible for the households to bid for specific properties, although the success of the bid was not guaranteed. In 2002, only 15% of the districts participated in CBL. This share was just above 20% in 2004 (DCLG, 2010). In our case, CBL schemes could only affect those who applied for social housing between late 2002 and early 2004 in the restricted group of districts participating in the scheme. Therefore, CBL schemes are unlikely to have substantially affected the randomness of the residential allocation of our 'social housing' subsample. 47 The results in the first and the second column of the bottom panel of Table 10 are derived from estimating our main specification on the social housing subsample. District ethnic diversity and segregation are likely to be less endogenous for this subsample than for the full estimation sample. This is because, in 2004, households who wanted to relocate were pushed to the back of the housing waiting list in the new district. This discouraged endogenous sorting into districts. Therefore, these results, in line with those obtained in Sections 4 and 5, can be seen as a further robustness check for our main specification.

Supply Side: the Provision of Social Activities
Our explanatory model abstracts from the supply side effects of district ethnic composition, i.e., how district ethnic composition can affect social participation through the provision of local public goods. In this section we discuss these supply side effects and we provide evidence that district ethnic segregation may incentivise PA through increased or better targeted provision of local sport facilities. The literature suggests that district ethnic diversity discourages the provision of local public goods, while district ethnic segregation encourages it (Alesina et al., 1999;La Ferrara and Mele, 2007). If local public goods include parks, sport facilities and clubs, district ethnic diversity (segregation) may decrease (increase) HA and PA. Our empirical model captures the total effect of district ethnic composition, i.e., the combination of the effects stemming from the demand and the supply side. We do not have an identification strategy to separate the demand and the supply effect. However, it is still interesting to discuss the correlation between district ethnic diversity and segregation and the provision of local public goods.
We are not aware of the existence of comprehensive data on provision of local public goods in England. However, we have information on the satisfaction with local sport facilities. As sport is an important component of PA, this could shed light on how district ethnic diversity and segregation correlate with the supply of PA. We use the 2005/2006 wave of the Active People Survey (APS): the closest in time to the LSYPE data used in our main analysis. APS contains information on around 360,000 respondents, including an indication of the district respondents live in and a question on satisfaction with local sport facilities.
We measure satisfaction with local sport facilities through an indicator equal to one if the respondent is fairly/very satisfied with local sport provision. 48 We collapse this variable at the district level and we regress it on the indices of district ethnic diversity and segregation plus the district level controls used to estimate the RS model in Table 6. Table 11 shows satisfaction with sport facilities is negatively correlated with district ethnic diversity and positively correlated with district ethnic segregation.
We are more confident about making causal statements about the results on district ethnic segregation in Table 11 than about the results on district ethnic diversity in the same table. Omitted deprivation positively correlated with district ethnic diversity and segregation and negatively correlated with the provision of sport facilities may lead to overestimating the negative effect of district ethnic diversity and underestimating the positive effect of district ethnic segregation. Therefore, we believe the results in Table 11 provide some evidence that district ethnic segregation may incentivise participation in PA through an increased or better targeted provision of local public goods.

Conclusions
This article studies the effect of district ethnic diversity and segregation on social participation. We focus on social participation for young people, as evidence suggests relational skills are formed during adolescence. We distinguish 'purposeful activities', which have been found to be beneficial for young people's development, from 'hanging around with friends', which we show to be associated with risky behaviours. We study the case of England, where improving the provision of purposeful activities has become central in the policy agenda (DfES, 2006).
We account for the potential endogeneity of district ethnic diversity and segregation in a number of different ways, including by proposing a novel set of instruments. All the methods that we use lead to the same conclusions. District ethnic diversity does not discourage participation in purposeful activities, but it does discourage hanging around with friends. The former is in line with Letki (2008); the latter is in line with the literature finding negative effects of ethnic diversity on social interaction (e.g., La Ferrara, 2000, 2002;Costa and Kahn, 2003;Pennant, 2005;Anderson and Paskeviciute, 2006;Charles and Kline, 2006;Putnam, 2007). District ethnic segregation increases hanging around with friends. Positive effects of district ethnic segregation on participation in purposeful activities are weak and are likely to be driven by supply side effects. To the authors' knowledge, these findings on the effect of district ethnic segregation on social participation are new.
Our findings have important policy implications. At least for the case of England, we show that ethnic diversity does not seem detrimental for young people's social participation. In fact, greater ethnic diversity could limit hanging around with friends and the potential risks associated with it. In contrast, ethnic segregation may be detrimental, as it is found to encourage hanging around. Thus, desegregation policies, in combination with policies improving the supply of purposeful activities, are more likely to have beneficial effects on young people's well-being than policies limiting migration to preserve ethnic homogeneity.
Our analysis focuses on the impact of ethnic diversity and segregation on social participation. Other forms of diversity and segregation, for example regarding income, social status, education, are likely to play a role in determining social participation. Our explanatory model can be adapted to analyse other forms of diversity and segregation. This is a fascinating exercise, which is left for future research.

A1. Introduction
This appendix describes young people's social participation. The first part describes the phenomenon of hanging around with friends in England. It answers four questions: (i) is hanging around near home pervasive among teenagers?; (ii) when do teenagers hang around?; (iii) is hanging around a problem?; (iv) what are the consequences of teenagers hanging around on adults' well-being? This analysis gives an idea of why hanging around has received so much attention in the policy debate and why a study of the phenomenon is needed. The second part investigates the short-and long-term consequences of hanging around and purposeful activities on teenagers' human capital accumulation, involvement in risky behaviours and social ties. We conclude social participation in teenage years has potentially long-lasting effects on crucial domains of people's lives. The third part presents evidence that young people prefer to interact with others from the same ethnic group. This is a crucial assumption in our explanatory model and thus it is important to show that it has some empirical support in our case.

A2. Hanging Around with Friends: A Description
This section describes what it means for teenagers to hang around with friends. We use the the British Crime Survey (BCS), which collects data on major crimes, as well as (perceived) antisocial behaviours, such as teenagers hanging around, noise and loud parties, vandalism and graffiti, drug dealing and using, drunkenness and alcohol-related violence. We use data from the 2004-2005 BCS adult main sample (over 45,000 respondents) and data on young people aged 14 to 15 from the 2009 under-16 sample (over 3,600 respondents). We restrict the analysis to young people aged 14 to 15: the age group we study in the body of the article. 49 Hanging around near home is a pervasive phenomenon among young people. Figure A1.a shows that almost 60% of the 14/15-year-old 2009 BCS respondents hang around at least once a week, and more than one in four respondents (27.41%) hangs around more than three times a week. In contrast, only 1 in 10 respondents (10.52%) never hangs around. Young people tend to hang around locally. When asked to report where they hang around (see Figure A1.b), respondents mention local shops in almost 6% of the cases and streets outside home in almost 13%. Some of the green areas, mentioned in around 38% of the cases, are also likely to be local. In sum, Figure A1   take place only at weekends, when teenagers have more free time. An overwhelming majority of BCS adult respondents (84.24%) report seeing teenagers hanging around during both weekends and weekdays (see Figure A2.a). Hanging around is mainly concentrated in the evening from 6 p.m. to 11 p.m. (see Figure A2.b). Hanging around in the morning is recorded in 1.12% of the  cases. This share reaches 8.32% if 'hanging around all the time' is added. Hanging around in the afternoon before 6 p.m. and in the night after 11 p.m. accounts for around 8% and 14.08% of the cases, respectively. In sum, the data suggest hanging around does not generally happen during school time (mornings and early afternoon), although the share of teenagers observed to hang around during school time is not negligible. Hanging around also takes place at night. This is potentially harmful for teenagers, as it interferes with their sleep. Figure A3 shows teenagers hanging around is perceived as a pervasive problem. Figure A3.a shows that around 20% of the young people consider teenagers hanging around a 'very big' or a 'fairly big' problem. Among adults, this figure reaches 29.56% (see Figure A3.b). Problems of comparable magnitude to teenagers hanging around, as reported by adults, are: rubbish lying around (very or fairly big problem for 29.29% of the respondents), vandalism and graffiti (very or fairly big problem for 27.26% of the respondents), people using or dealing drugs (very or fairly big problem for 24.98% of the respondents). Less than 22% of the adults consider drunk people to be a very or fairly big problem. Abandoned and burnt-out cars and racially motivated attacks are very or fairly big problems for only 11.25% and 6.23% of the adults, respectively. 50 Similar conclusions can be derived from Figure A4. Adults consider teenagers' hanging around the main problem of their neighbourhood both when they report the three main problems (Figure A4.a) and when they select the main one ( Figure A4.b). 50 To ensure comparability, the figures on the perception of different antisocial behaviour are derived on the same subset of respondents with valid data for all antisocial behaviours. Note that data in figures A3 and A4 are likely to measure a combination of how worrisome and how prevalent a behaviour is perceived.  Graph a: other is constructed as follows: people being attacked due to ethnicity 0.78%; people begging: 1.06%; people being insulted: 2.13%; abandoned or burnt-out cars: 2.75%; noisy neighbours or loud parties 4.00%. Graph b: other is constructed as follows: people being attacked due to ethnicity 0.80%; people begging: 0.84%; people being insulted: 1.42%; abandoned or burnt out cars: 1.84%; noisy neighbours or loud parties 4.63%; cannot choose one 2.03%. Figure A5 sheds light on why teenagers hanging around is perceived so negatively. Almost 65% of the adults consider teenagers hanging around on the streets to be deliberately antisocial. The antisocial behaviours most frequently reported are generally minor and include: swearing (15.3%), generally being a nuisance (14.36%), and being loud and noisy (13.93%). Drinking is the most frequently mentioned risky behaviour (11.22%). Other risky or illegal behaviours such as mugging, robbing, intimidating, threatening or physically assaulting people, damaging property or cars, fighting, doing graffiti, or taking drugs are infrequent individually, but all together account for 13% of the cases (see Figure A5.b). These data suggest that hanging around, while not criminal per se, is associated with illegal and antisocial behaviours. 51 Teenagers hanging around affect adults' well-being. Adults evaluated the effect on teenagers hanging around on the quality of their life using a 10-point scale, where 10 means 'very affected'. In 2004, the mean was 3.64 (for comparison, the means for vandalism and for people using or dealing drugs were respectively 3.34 and 3.10). The most common emotional response triggered by teenagers hanging around was annoyance (almost 30% of the cases, as shown in Figure A6). Many respondents also reported anger (14.10%), frustration (13.29%), worry (12.70%). More extreme expressions of distress are less frequent (crying: 0.49%; depression: 1.08%; anxiety/panic attacks: 1.37%; shock: 1.51%; stress: 4.58%), but all together account for more than 9% of the cases. Around 70% of the respondents reported direct consequences of the emotional distress caused by the teenagers hanging around. These include: avoiding certain places in the local area (13.34%), avoiding going out after dark (12.90%) or alone (8.32%), improving home or car security (7.65%), encouraging family or friends not to go out alone (7.12%), thinking about moving away (6.66%), using the car instead of walking (5.68%). Only 34% of the respondents made a complaint, but about one-third of them (11.13% of the total) complained directly to the police (see Figure A7.a). Interestingly, 16.7% of those who did not complain, failed to do it for 51 The under 16 module does not contain self-reported data on illegal behaviours taking place while young people hang around. However, it contains data on the reasons why adults (mainly the police or members of the public) have asked young people to move from the place where they were hanging around. These data are in line with those from the adult module, and show the young people are generally asked to move due to minor antisocial behaviours (e.g., being noisy), but also that behaviours such as smoking, drinking, intimidating others and damaging properties or areas are associated with hanging around. The question is only asked to those who did not complain to anyone. N = 2,301. Other is constructed as follows: dislike/fear of police/authorities: 0.31%; previous bad experience of the police 0.80%; just accept it 0.84%; private/personal/family matter 1.84%; waste of time 1.99%; problem is already known 2.37%; dealt with matter myself/ourselves 72%; other 2.14%. fear of reprisal by offenders. This suggests the relative low incidence of complaints is unlikely to be due to a low perceived salience of the problem.

A3. The Impact of Hanging Around with Friends and Purposeful Activities on Young People's Well-Being
To assess the effect of hanging around on teenagers' contemporaneous and future well-being, we use LSYPE data (and its longitudinal dimension), together with propensity score techniques. Hanging around with friends is measured at wave one. As outcome variables, we use contemporaneous (at wave one) and future (at waves five, six and seven) indicators of human capital investment and risky behaviours, and future indicators (at wave seven) of strength of social ties (see Table A1). 52 The analysis investigates whether HA is as negative as described in the BCS data 52 The wording of the questions used to derive the variables in Table A1 is the following: 1. 'During an average week in term time, on how many evenings do you do any homework?' (respondents are prompted to consider only weekdays from Monday to Friday inclusive); 2. 'How often do you read books, magazines or newspapers for pleasure?' (available answers: most days, more than once a week, once a week, less than once a week, never); 3. 'Have you applied for a place on a university course which will start either this year that is in September/October 2008 or next year, that is in September/October 2009?'; 4. 'Have you received any offers of places yet, either conditional offers or unconditional offers?'; 7. 'Do you ever smoke cigarettes at all?'; 8. 'Have you ever had a proper alcoholic drink? That is a whole drink, not just a sip. Please do not count drinks labelled low alcohol'; 9. 'Have you ever tried cannabis even if only once?'; 10. 'Have you ever written on walls with spray cans?'; 11. 'Have you ever smashed, slashed or damaged public property or something in a public place?'; 12. 'Have you ever taken something from a shop, supermarket, or department store without paying?' 13. 'Have you ever taken part in fighting or some sort of disturbance in public for example, at a football ground, a railway station, music festival, riot, demonstration or just in the street?'; 14. 'Have you ever had sex without using precautions or contraception?' (respondents are prompted not to include any times when trying for a baby); 15. the economic journal [august presented above. We investigate whether HA displaces investment in human capital (reading or study time) and encourages risky behaviours, and whether these effects persist over time. We also investigate whether HA engenders the creation of long-lasting friendship ties. This potentially positive effect of HA is overlooked in the political debate. Table A2 compares the means of the outcome variables for respondents who hang around at wave one (HA respondents) and respondents who do not hang around at wave one (non-HA respondents). All differences are statistically significant. HA respondents experience lower human capital accumulation than non-HA respondents. For example, at age 14/15, HA respondents spend 14% less time (almost half the evening) doing homework. By age 18/19, HA respondents are over 25% (12 p.p.) less likely to apply to and receive offers from higher education institutions. 'Have you ever taken cannabis?'; 16. 'Have you ever taken other drugs?'; 17 'How many close friends do you have, that is friends you could talk to if you were in some sort of trouble?' (available answers are: none, 1, 2-3, 4-5, 6-9, 10 or more). Variables in rows 5 and 6 are derived from a number of different questions.  Two years later, they are 27% (11 p.p.) less likely to be in higher education and 31% (3 p.p.) less likely to be in a Russell group university. 53 HA respondents are also more likely than non-HA respondents to be involved in risky behaviours. At the age of 14/15, HA respondents are 69% (23 p.p.) more likely to have tried alcohol and 81% (7 p.p.) more likely to have shoplifted. They are more than twice as likely to have tried cigarettes or cannabis, and to have been involved in graffiti, vandalising, and fights. These differences in risky behaviours reduce only slightly with time. By age 19/20, HA respondents are 61% (24 p.p.) more likely to have had unsafe sex, 67% (15 p.p.) more likely to have tried cannabis, and 86% (6 p.p.) more likely to have tried other drugs. Finally, Table A2 shows that HA respondents have 6% more close friends in adulthood. The differences in means in Table A2 cannot be interpreted as causal. They are likely to be due to a combination of selection and causal effect. Young people who are less interested in studying and reading and more attracted by risky behaviours are likely to self select into HA. This is the selection effect. The potential causal effect is three-fold. First, the time spent hanging around is subtracted to studying and reading. Thus, HA may harm human capital accumulation through a decrease in study/reading time. Second, hanging around takes place without the supervision of adults. This reduces the likelihood of being sanctioned and thus the cost of risky behaviours. 53 The Russell Group is a self selected group of 24 public funded universities. Affiliation to the Russell group is often perceived as a signal of high quality education.  Third, if 'risky' young people self-select into HA, peer effects may encourage less study and more risky behaviours. The fact that HA respondents form stronger friendship ties, while positive per se, might exacerbate the role of negative peer effects.
To shed light on the causal effect of hanging around (Average Treatment Effect on the Treated (ATET)) on young people's human capital accumulation, risky behaviours and social ties, we use propensity score techniques. We use the distance-weighted radius matching with bias adjustment by Lechner et al. (2011), which weights the control observations using the inverse of their C 2019 Royal Economic Society.  Notes: The grey dots indicate the bias in the covariates before the matching, the black 'x' signs indicate the bias in the covariates after the matching. The bias after the matching is computed after weighting the control observations using the inverse of their distance from the corresponding treated observation.

the economic journal
[august distance from the corresponding treated observation. 54 Distance-weighted radius matching limits the likelihood of a bad match, and thus is more precise and less biased than nearest neighbour matching. This is particularly desirable in our case, where the treatment group is larger than the control group. Moreover, the estimator has been proven to be robust to misspecification of the propensity score (Huber et al., 2013(Huber et al., , 2015. With propensity scores, identification of ATET relies on having observable covariates that permit to predict the treatment such that, conditional on the propensity score, the potential outcomes in the absence of treatment are independent of the treatment assignment (Conditional Independence Assumption CIA). To try to satisfy this assumption, we estimate the propensity score using a large set of individual, household and neighbourhood characteristics from wave one, including a set of characteristics measured at birth and in early age. 55 In addition, identification of ATET relies on each treated observation having a non-treated counterpart with similar characteristics and thus propensity score (common support assumption).
For each outcome analysed, Figure A8 shows the distribution of the propensity score for the treated and the control observations before and after the matching. The post-matching distributions are derived by re-weighting the pre-matching propensity scores using weights obtained in the matching procedure. The post-matching distributions for the treated and the control observations overlap for all outcomes. 56 This shows the common support assumption holds. While the pre-matching distributions of the propensity score for the treated and the control observations are clearly different, the post-matching distributions are almost identical. This provides a first glance into the ability of the propensity scores of making treated (HA respondents) and control (non-HA respondents) observations comparable. Figure A9 shows the standardised percentage bias (i.e., covariate imbalance) for each covariate before and after the matching. While the pre-matching bias is substantial (between −40 and +40 standardised percentage bias), the post-matching bias is close to zero for all covariates. Diagnostics on overall post-matching covariates imbalance are reported in Tables A3 and A4. 57 All tests suggest our propensity scores are satisfactory. The R statistics are close to one, the B statistics are generally below 16 (see Table A3). The mean and the median bias, that before the propensity score exceeded 5% and 7%, respectively, now drop at or below 1% (see Table A4). The pseudo R 2 are very close to zero (and much smaller than those before the matching), and 54 The weights are then used in a regression to correct the bias due to the mismatch. The estimator is implemented in stata through the command 'radiusmatch' (Huber et al., 2012). 55 These variables are: respondent's characteristics, i.e., gender, ethnicity, birth weight, whether the respondent was born early, whether the respondent has attended a nursery in her early childhood, month of birth (in months of distance from the previous August, when school starts); household characteristics, i.e., household income (in quintiles), whether the household has a computer at home, whether the household has internet access, whether the household has family dinners every or most days, whether the main parent is female, mother's and father's education (degree, above GCSE, GCSE and below, no education), mother's and father's employment status (full time, part time, not working), religion of main parent, health of main and second parent, number of cars (none, one, two, more than two), whether mother/father was NEET when the respondent was five, total months of unemployment of mother/father since the respondent was born, whether the respondent has ever lived in a single-parent household; neighbourhood characteristics, i.e., indices of ethnic fractionalisation and segregation, as used in the paper. We also add survey weight as in Rubin (2001). Dummy variables are used with categorical variables, with a separate dummy indicating missing data. 56 Weights for the treated observations are equal to 1. The number of off-support observations ranges from a minimum of zero to a maximum of 14. These observations have not been used in the analysis. 57 R is the ratio between the variance of the propensity score index for HA and non-HA respondents; B is the standardised difference in the means of the propensity scores computed for the HA and the matched non-HA respondents.
To be satisfactory, R should be close to one, and surely between 0.5 and 2, while B must be below 25 (Rubin, 2001). The pseudo R 2 is derived from regressing the propensity score on the covariates used to estimate the propensity score.  Notes: LSYPE waves 1, 5, 6, 7. ATET are estimated using the radiusmatch stata command (Huber et al., 2012). the likelihood test is not able to reject the hypothesis of joint insignificance of the observables in explaining the propensity score (see Table A4). Table A5 shows the ATETs estimated via propensity score matching. The effect of HA suggested by the ATETs are all statistically significant (generally at the 1% level), and maintain the same sign as the t-tests in Table A2. However, the ATETs are smaller in magnitude. They suggest HA respondents spend 7% less time doing homework (one-fifth of a day) than non-HA respondents and are 4% (3 p.p.) less likely to read for pleasure at least once a week. HA respondents are 10% and 12% (6 and 5 p.p.) less likely to apply to and receive offers from higher education institutions, respectively. They are also 13% (5 p.p.) less likely to be in higher education at age 20/21. The ATET for attending a Russell group university is −12% (1 p.p.), but it is at the verge of statistical significance.
The estimated ATETs for risky behaviours at wave one are between five (done graffiti) and 13 p.p. (tried alcohol). These effects translate into 112% and 39%, respectively. The ATETs for having had unsafe sex (17 p.p.) and having tried cannabis by age 19/20 (12 p.p.) are sizeable. In percentage terms, these ATETs suggest that HA increases the probability of both risky behaviours by around a half. Equally, the ATET for trying other drugs, only 4 p.p., signals an increase of more than half, when seen in percentage. Finally, the ATET for the number of close friends at age 19/20 is small in magnitude (around 5%), and very close to the estimate obtained using simple t-tests and reported in Table A2.
In summary, our results suggest the differences observed in the mean outcomes of HA and non-HA respondents are partially due to selection. However, there might be a causal effect of C 2019 Royal Economic Society. the economic journal [august Notes: LSYPE waves 1, 5, 6, 7. ATET are estimated using the radiusmatch stata command (Huber et al., 2012).
HA. In the case of the human capital outcomes, the ATETs estimated through propensity score matching are about half the size of those suggested by t-tests. Such a drop in the estimated effect following the use of propensity score matching is not observed in the case of risky behaviours and social ties (compare, for example, the results in Tables A2 and A5 for outcomes 10 and 17). This suggests that the negative effect of HA on risky behaviours and the (small) positive effect on social ties is more likely to be causal than the one on education. The results on risky behaviours are in line with our analysis of BCS data: they suggest risky behaviours do take place when young people hang around. Moreover, HA is likely to affect risky behaviours through peer effects, as vandalism, fights and drug initiation are more likely to be group activities than studying. The results on social ties are also likely to be at least partially causal, as spending time hanging around with friends may strengthen friendship ties. We now look at the effects of purposeful activities (PA). Table A6 compares the means of the outcomes in Table A1 for those who take part in PA at wave one (PA respondents) and those who do not take part in PA at wave one (non-PA respondents). PA is associated with higher human capital accumulation. For example, compared with non-PA respondents, PA respondents spend around one-fifth of an evening more doing homework at age 14/15 and are 8% more likely to be in higher education at age 20/21. The picture on risky behaviour is more nuanced. Notes: LSYPE waves 1, 5, 6, 7. ATET are estimated using the radiusmatch stata command (Huber et al., 2012). drugs). Finally, at age 19/20 PA respondents have on average almost one close friend more than non-PA respondents. Again, the differences in Table A6 cannot be considered causal. The potential causal effect of PA is unclear a priori. On the one hand, PA foster skills like grit, self-confidence, trust and cooperation. These skills are likely to boost human capital accumulation and discourage risky behaviours. On the other hand, PA subtract time to homework and reading and put young people in contact with other people of the same age. Therefore, PA may interfere with human capital accumulation and expose young people to peer effects encouraging risky behaviours.
As with HA, we use propensity score matching to shed light on the causal effect of PA. 58 Figures A10 and A11, and Tables A7 and A8 suggest the matching has successfully achieved comparability between the treatment and the control group. Figure A10 shows the propensity scores for treated and controls have a large common support. Moreover, it shows the distributions of the propensity score computed for treated and control observations are initially quite different, but then become indistinguishable when reweighed. Figure A11 shows the standardised bias for the covariates is initially sizeable, but it dramatically reduces after the matching. Finally, all measures for overall covariate imbalance are satisfactory (see Tables A7 and A8).
The ATETs for PA obtained via propensity score matching are shown in Table A9. The results on human capital accumulation and social ties maintain the same sign-and roughly the same size-as the t-tests in Table A6. For example, the ATETs suggest PA respondents are likely to the economic journal [august spend over one-fifth of an evening doing homework more than non-PA respondents. Moreover, PA respondents are 19% (almost 7 p.p.) more likely to apply to university than non-PA respondents at age 18/19, and 21% more likely (over 6 p.p.) to be in higher education two years later. The ATETs for risky behaviours suggest at age 14/15 PA respondents are 20% less likely (around 2 p.p.) to have tried cigarettes and cannabis, but also 30% more likely (5 p.p.) to be involved in fights. 59 For risky behaviours the ATETs suggest a more desirable effect of PA than the t-tests, as most of the positive signs estimated through t-tests become zeros or turn negative when propensity scores are used.
In summary, we find that PA are associated with-and possibly determine-higher human capital accumulation and greater number of close friends in adulthood. The association between PA and risky behaviours is less clear. Our analysis suggests PA participants are more likely to select into risky behaviours, but the causal effect of PA may go in the opposite direction, and PA may prevent young people from being involved in (at least some) risky behaviours.
Our findings on HA are entirely new, as no literature investigates this pervasive and policyrelevant behaviour. Our findings on PA are in line with the existing literature that generally finds positive effects of PA on educational attainment (Long and Caudill, 1991;Barron et al., 2000;Eide and Ronan, 2001;Lipscomb, 2007;Lechner, 2009;Pfeifer and Cornelißen, 2010;Rees and 59 The result on fights may be partially due to the fact that LSYPE reports as fights also those occurred in political rallies, and participating in political rallies is a component of PA.  where the control observations within a certain radius from the treated have been weighted proportionally to the inverse of their distance from the treated. Weights for the treated observations are equal to 1. Sabia, 2010;Stevenson, 2010;Felfe et al., 2016) but no effects on antisocial behaviours (Felfe et al., 2016). 60 The results in this appendix rest on the strong assumption of selection on observables, and thus one needs to be cautious in claiming causality. 61 However, our propensity score analysis is 60 Negative effects of sport participation on crime are found by Caruso (2011), but the identification of the effect only relies on regional variation for Italian regions. 61 We are fairly confident about the robustness of our propensity score analysis for two reasons. First, our propensity scores are based on a large set of variables. This reduces any unobservable effect potentially left in the error term. Second, C 2019 Royal Economic Society.   Notes: LSYPE waves 1, 5, 6, 7. ATET are estimated using the radiusmatch stata command (Huber et al., 2012).
not able to revert the initial conclusions based on raw data. We still conclude HA is negatively associated with human capital accumulation and positively associated with risky behaviours. In contrast, PA are positively associated with accumulation of human capital. This suggests that there is scope for economists to work on the identification, under weaker assumptions, of the causal mechanisms linking HA and PA to these crucial outcomes. Also, this suggests the importance of studying the determinants of HA and PA, as the body of the article does.

A4. Social Interaction and Preference for Own Ethnic Group.
In this section, we use LSYPE data to provide empirical support to our hypothesis that young people prefer to interact with other young people from their own ethnic group. For each of the main ethnic groups (White, Indian, Pakistani, Bangladeshi, Black Caribbean, Black African), Table A10 reports: (i) the share of that ethnic group in the sample (as the data are weighted, this is an estimate of the share in the population); (ii) information on the ethnicity of the friends the selection on observables, which can be inferred by the difference between the t-tests and the ATETs are tiny, and so is likely to be the selection on unobservables. This argument is based on the assumption that the bias due to observables has the same sign as the bias due to unobservables, and the latter is at most as important as the former (Altonji et al., 2005).
respondents from that ethnic group have at school; (iii) information on the ethnicity of the friends respondents from that ethnic group have outside school. 62 Table A10 suggests that young people form ethnically homogeneous groups of friends and do it by choice. Consider the case of Pakistanis (third panel from the top), who are around 2% of the total. If friends' ethnicity was random, Pakistanis should have around 2% of Pakistani friends and thus report very few friends who are Pakistani themselves. However, 68.6% of the Pakistanis say at least half of their friends at school are Pakistani. This percentage reaches 79.12% in the case of friends outside school. This suggests that groups of friends are more homogeneous than if they were formed randomly. Table A10 also shows that friendship ties outside school are more ethnically homogeneous than those formed at school. While the ethnic mix of friends from school is driven by the ethnic composition of the school catchment areas, friends outside schools are more likely to be freely chosen. This suggests that the ethnic homogeneity of friendship groups is at least partially driven by preferences. If young people prefer to form ethnically homogeneous groups, HA and PA are likely to be affected by district ethnic diversity and segregation.

2019]
ethnic composition and participation 2505 Notes: LSYPE waves 1, 5, 6, 7. ATET are estimated using the radiusmatch stata command (Huber et al., 2012). Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Individual controls: Male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Weighted data. Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Reference categories: female, born in 1990, White British, bottom income quintile, mother tongue is English, mother with no education, mother younger than 35 years old, mother works full time. Weighted data. Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Weighted data. Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Reference categories: female, born in 1990, White British, bottom income quintile, mother tongue is English, mother with no education, mother younger than 35 years old, mother works full time. Weighted data. Notes: Tests are computed using the ivreg2 stata command by Baum et al. (2002). Instruments IV2: proximity to the closest and the second closest port of entry and its higher terms (up to the squares), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest and the second closest port and the number of ports within 20, 50, 100 km. Instruments IV3: distances from all the ports of entry and mill towns. Instruments IV4: binary variables capturing the number of ports (mill towns) within a 20, 50,100 km radius. Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Results are computed using the ivreg2 stata command by Baum et al. (2002). Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Instruments IV2: proximity to the closest and the second closest port of entry and its higher terms (up to the squares), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest and the second closed port and the number of ports within 20, 50, 100 km. Instruments IV3: distances from all the ports of entry and mill towns. Notes: Tests are computed using the ivreg2 stata command by Baum et al. (2002). Instruments IV1: proximity to the closest port of entry and its higher terms (up to the third), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest port and the number of ports within 20, 50, 100 km. Notes: Tests are computed using the ivreg2 stata command by Baum et al. (2002). Instruments IV1: proximity to the closest port of entry and its higher terms (up to the third), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest port and the number of ports within 20, 50, 100 km.

2019]
ethnic composition and participation 2515 Notes: Standard errors in parentheses, *p < 0.1, **p < 0.05, ***p < 0.01. Results are computed using the ivreg2 stata command by Baum et al. (2002). Individual controls: male, born in 1989, ethnicity dummies. Family controls: income quintiles, English not mother tongue, main parent is female, mother's education, age and employment status. Weather controls: total sunshine duration, total rainfall. Each LSYPE respondent is assigned the weather information collected from the closest meteo station in the interview month. Instruments IV1: proximity to the closest port of entry and its higher terms (up to the third), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest port and the number of ports within 20, 50, 100 km. Weighted data. Notes: Tests are computed using the ivreg2 stata command by Baum et al (2002). Instruments IV1: proximity to the closest port of entry and its higher terms (up to the third), number of ports in a radius of 20, 50, 100 km, and interactions between the proximity to the closest port and the number of ports within 20, 50, 100 km.
Greens in ward 1 and Purples in ward 2 have now increased incentives to do HA. Purples in ward 1 and Greens in ward 2 have now decreased incentives to do HA. Asn g1 + n + n p1 > n g1 − n + n p1 , ∀n = 1, . . . ,n g1 − 1, at the aggregate level, this translates into an increase in the total incentives of doing HA. Since district ethnic diversity remains unaffected: ≤ ∂EU(PA) ∂n g1 . Sincē n g1 + n + n p1 >n g1 − n + n p1 , ∀n = 1, . . . ,n g1 − 1, more people have incentive to switch from PA to HA. (b), (c). For a given N g we need to find the values n g1 and n g2 that keep district ethnic segregation equal toD ∀N g . Since n g2 = N g − n g1 , we need to find n g1 that solves: The two solutions of (C1) are: n g1 = 1 2 +D N g and consequently n g2 = 1 2 −D N g , (C2) n g1 = 1 2 −D N g and consequently n g2 = 1 2 +D N g .
The results (C2) and (C3) are symmetric and thus we focus on (C2) only, where n g1 ≥ n g2 .