The Stockholm University Linnaeus Center for Integration Studies (SULCIS)

On Fragile Grounds A replication of ldquoAre Muslim immigrants different in terms of cultural integrationrdquo

Technical documentation

Mahmood Arai Jonas Karlsson and Michael Lundholm

Working Paper 20092 ISSN 1654-1189

On Fragile Grounds

A replication of Are Muslim immigrants

different in terms of cultural

integration

Technical documentation

Mahmood Arailowast Jonas Karlssondagger and Michael LundholmDagger

December 19 2008

Abstract

This is a technical documentation of Arai et al (2008) which repli-cates ldquoAre Muslim Immigrants Different in terms of Cultural Integra-tionrdquo by Alberto Bisin Eleonora Patacchini Thierry Verdier andYves Zenou published in Journal of European Economic Association6 445-456 2008

Bisin et al (2008) report that they have 5963 observations in theirstudy Using their empirical setup we can only identify 1901 relevantobservations in the original data After removing missing values we areleft with 818 observations We cannot replicate any of their resultsand our estimations yield no support for their claims

lowastCorresponding author Department of Economics and SULCIS Stockholm UniversitySE 106 91 Stockholm Sweden mahmoodarainesuse

daggerThe Institute for Social Research and SULCIS Stockholm Universityjonaskarlssonsofisuse

DaggerDepartment of Economics Stockholm University michaellundholmnesuse

1

On Fragile Grounds

A replication of Are Muslim immigrants

different in terms of cultural

integration

Technical documentation

Mahmood Arailowast Jonas Karlssondagger and Michael LundholmDagger

December 19 2008

Abstract

This is a technical documentation of Arai et al (2008) which repli-cates ldquoAre Muslim Immigrants Different in terms of Cultural Integra-tionrdquo by Alberto Bisin Eleonora Patacchini Thierry Verdier andYves Zenou published in Journal of European Economic Association6 445-456 2008

Bisin et al (2008) report that they have 5963 observations in theirstudy Using their empirical setup we can only identify 1901 relevantobservations in the original data After removing missing values we areleft with 818 observations We cannot replicate any of their resultsand our estimations yield no support for their claims

lowastCorresponding author Department of Economics and SULCIS Stockholm UniversitySE 106 91 Stockholm Sweden mahmoodarainesuse

daggerThe Institute for Social Research and SULCIS Stockholm Universityjonaskarlssonsofisuse

DaggerDepartment of Economics Stockholm University michaellundholmnesuse

1

Contents

1 Introduction 3

2 Data and variable description 4

21 Data 422 Reading data and selecting variables 523 Defining the relevant (ethnic minority) sample 524 Recoding of missing values 625 Variable definitions 6

251 Religious affiliation 7252 Importance of religion 7253 Attitude towards inter-marriage 7254 Importance of racial composition in schools 8255 Born in the UK 8256 Age at and years since arrival 8257 Female 9258 Arranged marriage 9259 Discrimination 92510 Children 102511 No British education 112512 British basic education 112513 British higher education 122514 Foreign education 122515 Labour market status 122516 No parents 152517 Contacts with parents 152518 English language 152519 Household income 162520 Ward variables 17

26 Discrimination own ethnicity 1827 Defining the subset 1828 Sample statistics 18

3 Regression Results 20

4 Concluding remarks 21

5 Production notes 21

2

1 Introduction

This is a replication of the empirical results reported in Bisin et al (2008)They use British data and analyse how Muslims and non-Muslims differin cultural integration measured as (i) Importance of Religion (ii) Atti-tude Towards Inter Marriage and (iii) Importance of Racial Composition inSchools1

In the abstract of their paper they write

ldquo Muslims integrate less and more slowly than non-Muslims We also find no evidence that segregated neighbourhoods breedintense religious and cultural identities for ethnic minorities es-pecially for Muslimsrdquo (Bisin et al 2008 p 245)

We wanted to check the robustness of their results when considering theethnic and religious heterogeneity within both groups Muslims and non-Muslims Among other things we were concerned about the measures ofcultural values used in the paper These measures capture ethnic and re-ligious attributes in different degrees for different groups For example thevariable Attitudes towards Inter-Marriage with the majority UK populationcaptures only inter-ethnic marriage for the Christian ethnic minorities butboth inter-ethnic and inter-religious marriage for Muslims

However already an initial inspection of data disclosed that the numberof observations in Bisin et al (2008) exceeded the total number of observa-tions in the ethnic minority sample We communicated this to the authorsand they answered that there were some coding errors We have receivedrevised codes and a revised version of their specifications and tables Theirrevised codes yield fewer observations than the sample in the published ver-sion but still more than we can identify in the relevant sample of the originaldata As far as we can see a source of the large number of observations intheir revised codes is that dummy variable definitions include observationswith missing values in the reference categories (defined as zeros) The un-derlying codes to the published paper were however not made available andthe exact nature of the original errors are therefore unknown to us

Bisin et al (2008) report that they have 5963 observations in their studywhereas the ethnic minority sample in Berthoud et al (1997) consists of5226 observations Implementing their empirical setup we can only identify1901 relevant observations in the original data After removing missingvalues we are left with 818 observations Using the remaining sample andrunning their specifications we find no results that support their claims Ourreplication therefore stopped here and we did not perform any sensitivityanalysis The great loss of observations implies that the remaining sample

1To facilitate comparability we use the same labels on the variables as Bisin et al(2008)

3

is most likely not representative Therefore we hesitate to draw inferencefrom the regressions results

In this paper we only document the replication and report and commentresults using the variable definitions the variable names and the specifica-tions used in Bisin et al (2008) We choose a procedure that makes it easyto reproduce our results Influenced by Koenker and Zeileis (2007) we usean integrated approach where data management estimations and the textthat rely on these computations are all integrated in one single file Thisstrategy has the advantage that it makes is easy to adjust the codes andautomatically generate a revised version of the paper

All data analysis is made in R (R Development Core Team 2008) andall code files related to this project can be found on httppeoplesuse

~lundhfragile_groundsIn this technical documentation we present our results in greater detail

but also all our working procedures variables definitions etc In additionthe central part of our codes are included with typeset comments This isdone as an attempt to implement Literate Statistical Programming

The remainder of the paper is organised as follows The data and variabledefinitions are described in Section 2 Regression results are presented inSection 3 The paper is concluded in Section 4 Finally the productionprocedure is described in Section 5

2 Data and variable description

21 Data

The data set is the Fourth National Survey of Ethnic Minorities 1993-1994(FNSEM) see (Berthoud et al 1997)2 Itrsquos main objective were

ldquoto describe the social and economic conditions of Britainrsquosmain ethnic minority groups including their health and tocompare these with the social and economic conditions ofthe white majority

to assess changes over time through comparisons with otherwork

to show how the position of ethnic minority groups is relatedto the social and ethnic compositions of the areas in whichthey live

to explore diversity among different ethnic minority groups

2The data can be accessed from the UK Data Archive (UKDA) via Athens UKData Archive is found at httpwwwdata-archiveacuk and Athens at httpwwwathensacuk

4

1 Introduction

This is a replication of the empirical results reported in Bisin et al (2008)They use British data and analyse how Muslims and non-Muslims differin cultural integration measured as (i) Importance of Religion (ii) Atti-tude Towards Inter Marriage and (iii) Importance of Racial Composition inSchools1

In the abstract of their paper they write

ldquo Muslims integrate less and more slowly than non-Muslims We also find no evidence that segregated neighbourhoods breedintense religious and cultural identities for ethnic minorities es-pecially for Muslimsrdquo (Bisin et al 2008 p 245)

We wanted to check the robustness of their results when considering theethnic and religious heterogeneity within both groups Muslims and non-Muslims Among other things we were concerned about the measures ofcultural values used in the paper These measures capture ethnic and re-ligious attributes in different degrees for different groups For example thevariable Attitudes towards Inter-Marriage with the majority UK populationcaptures only inter-ethnic marriage for the Christian ethnic minorities butboth inter-ethnic and inter-religious marriage for Muslims

However already an initial inspection of data disclosed that the numberof observations in Bisin et al (2008) exceeded the total number of observa-tions in the ethnic minority sample We communicated this to the authorsand they answered that there were some coding errors We have receivedrevised codes and a revised version of their specifications and tables Theirrevised codes yield fewer observations than the sample in the published ver-sion but still more than we can identify in the relevant sample of the originaldata As far as we can see a source of the large number of observations intheir revised codes is that dummy variable definitions include observationswith missing values in the reference categories (defined as zeros) The un-derlying codes to the published paper were however not made available andthe exact nature of the original errors are therefore unknown to us

Bisin et al (2008) report that they have 5963 observations in their studywhereas the ethnic minority sample in Berthoud et al (1997) consists of5226 observations Implementing their empirical setup we can only identify1901 relevant observations in the original data After removing missingvalues we are left with 818 observations Using the remaining sample andrunning their specifications we find no results that support their claims Ourreplication therefore stopped here and we did not perform any sensitivityanalysis The great loss of observations implies that the remaining sample

1To facilitate comparability we use the same labels on the variables as Bisin et al(2008)

3

is most likely not representative Therefore we hesitate to draw inferencefrom the regressions results

In this paper we only document the replication and report and commentresults using the variable definitions the variable names and the specifica-tions used in Bisin et al (2008) We choose a procedure that makes it easyto reproduce our results Influenced by Koenker and Zeileis (2007) we usean integrated approach where data management estimations and the textthat rely on these computations are all integrated in one single file Thisstrategy has the advantage that it makes is easy to adjust the codes andautomatically generate a revised version of the paper

All data analysis is made in R (R Development Core Team 2008) andall code files related to this project can be found on httppeoplesuse

~lundhfragile_groundsIn this technical documentation we present our results in greater detail

but also all our working procedures variables definitions etc In additionthe central part of our codes are included with typeset comments This isdone as an attempt to implement Literate Statistical Programming

The remainder of the paper is organised as follows The data and variabledefinitions are described in Section 2 Regression results are presented inSection 3 The paper is concluded in Section 4 Finally the productionprocedure is described in Section 5

2 Data and variable description

21 Data

The data set is the Fourth National Survey of Ethnic Minorities 1993-1994(FNSEM) see (Berthoud et al 1997)2 Itrsquos main objective were

ldquoto describe the social and economic conditions of Britainrsquosmain ethnic minority groups including their health and tocompare these with the social and economic conditions ofthe white majority

to assess changes over time through comparisons with otherwork

to show how the position of ethnic minority groups is relatedto the social and ethnic compositions of the areas in whichthey live

to explore diversity among different ethnic minority groups

2The data can be accessed from the UK Data Archive (UKDA) via Athens UKData Archive is found at httpwwwdata-archiveacuk and Athens at httpwwwathensacuk

4

is most likely not representative Therefore we hesitate to draw inferencefrom the regressions results

In this paper we only document the replication and report and commentresults using the variable definitions the variable names and the specifica-tions used in Bisin et al (2008) We choose a procedure that makes it easyto reproduce our results Influenced by Koenker and Zeileis (2007) we usean integrated approach where data management estimations and the textthat rely on these computations are all integrated in one single file Thisstrategy has the advantage that it makes is easy to adjust the codes andautomatically generate a revised version of the paper

All data analysis is made in R (R Development Core Team 2008) andall code files related to this project can be found on httppeoplesuse

~lundhfragile_groundsIn this technical documentation we present our results in greater detail

but also all our working procedures variables definitions etc In additionthe central part of our codes are included with typeset comments This isdone as an attempt to implement Literate Statistical Programming

The remainder of the paper is organised as follows The data and variabledefinitions are described in Section 2 Regression results are presented inSection 3 The paper is concluded in Section 4 Finally the productionprocedure is described in Section 5

2 Data and variable description

21 Data

The data set is the Fourth National Survey of Ethnic Minorities 1993-1994(FNSEM) see (Berthoud et al 1997)2 Itrsquos main objective were

ldquoto describe the social and economic conditions of Britainrsquosmain ethnic minority groups including their health and tocompare these with the social and economic conditions ofthe white majority

to assess changes over time through comparisons with otherwork

to show how the position of ethnic minority groups is relatedto the social and ethnic compositions of the areas in whichthey live

to explore diversity among different ethnic minority groups

2The data can be accessed from the UK Data Archive (UKDA) via Athens UKData Archive is found at httpwwwdata-archiveacuk and Athens at httpwwwathensacuk

4

to describe perceptions and experience of racial discrimina-tion and social harassmentrdquo

Berthoud et al (1997)

For our coding we have used FNSEM (1993a) which contains the projectinstructions and FNSEM (1993b) which is the data description file includedin the files obtained when the entire data set is downloaded from the UKDA

In the following we present how the original data are used to define thedata set used in the estimations We present extracts of our R code (RDevelopment Core Team 2008) with extensive comments and discussionsFor details about our working procedures and how we document the researchsee section 5 on page 21 In the code chunks ldquogtrdquo denotes the R prompt andldquo+rdquo continuation of the previous line

22 Reading data and selecting variables

We load package foreign to read STATA data format Data is read from theunpacked Stata-version of the data and ldquo_rdquo in variable names are convertedto ldquordquo

gt library(foreign)

gt FNSEM lt- dataframe(readdta(3685dta

+ convertunderscore=TRUE))

After reading the data we select a subset of variables to be used Thiscode is in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

23 Defining the relevant (ethnic minority) sample

The data consist of two samples Ethnic Minorities and Whites We areonly interested in the former and remove all Whites The variable ethnic

indicates ethnic group of the individual according to the British standardand is used for this purpose

One of the three measures of cultural integration in Bisin et al (2008) isImportance of Religion Whether a respondent has a religion or belongs toa church is registered in question s6 Those who do not have a religion ordo not belong to a church are coded 2 we remove these observations fromthe sample since they cannot be classified in a religious group

A variable used in Bisin et al (2008) concerns the role of the respondentand his or her parents about choosing the respondentrsquos husbandwife Sincethis information is only available for married and previously married personsthe unmarried persons are removed from the sample

Furthermore respondents were faced with one out of three question-naires (green (catageory 1) yellow (category 2) and pink (category 3)) The

5

questions involved in the study are only answered by individuals who werefaced with the green questionnaire Therefore we keep only these in thesample

gt FNSEM lt- FNSEM[FNSEM$ethnic=white amp

+ isna(FNSEM$ethnic)]

gt FNSEM lt- FNSEM[FNSEM$s6=2]

gt FNSEM lt- FNSEM[FNSEM$a1e=3]

gt U lt- FNSEM lt- FNSEM[FNSEM$question==1]

An issue where Bisin et al (2008) is imprecise is whether the questionsthey address regard Muslimsnon-Muslims or Muslimnon-Muslim immi-grants Different sample selections are possible here The model specifi-cations in Bisin et al (2008) implies that White Muslims are excluded andnative ethnic minority Muslims are included This sample definition does notmatch Bisin et al (2008) writing using the terms Muslim and non-Muslimimmigrants as the sample includes natives

24 Recoding of missing values

The data set contains several codes for missing values These missing valuescan be of different characters eg nonndashavailables lsquocanrsquot sayrsquo or because therespondent was filtered in a previous filter question We employ the strategyto code all these as nonndashavailables in R ie code NA In the data set genuinenon-availables are generally coded as minus1 We set all minus1 to NA in the entiredata set

gt isna(U) lt- U==-1

In addition to minus1 several other codes (na 7 8 9 98 99 997 and 999)are occasionally used in the data set to indicate various unknown categoriesSome of minus1 and other unknown categories are nonndashavailables and have tobe deleted in estimations This is done after we have coded all our variablesin section 27 In questions following a filter question NA may have to beset in a category Some of these other codes used to denote unknowns aregenuine NArsquos and has to be removed Others will be included in a categoryThis is done variable by variable below

25 Variable definitions

Using the same variable names as in Bisin et al (2008) we define the vari-ables at the precision described by the authors We here give our interpre-tation of the variable definitions in Bisin et al (2008)

6

251 Religious affiliation

Question s6 asks whether the respondent belongs to a church or has a re-ligion For those who answer yes question s7 asks which that church orreligion is We define two religious affiliations muslim (which are all whoanswered category 3 (muslim) on question s7) and non-muslim (all who didnot answer category 3 (muslim) on question s7) All nonndashreligious that isthose who answered category 2 on question s6 are already removed from thesample Observations containing na are recoded to NA (FNSEM 1993a p112f)

gt isna(U$s7) lt- U$s7==na

gt U$Religion lt- ifelse(U$s7==muslimmuslimnon-muslim)

252 Importance of religion

Question s9 is about the importance of religion To grade the Importanceof Religion respondents have to choose between the following categories 1not at all important 2 not very important 3 fairly important or 4 veryimportant Following standard coding practice of such questions 1 and 2should be one category and 3 and 4 another but Bisin et al (2008) chooseto put 1 2 and 3 in the same category

Among those who have answered the question very few have chosenthe alternative 1 or 2 in their answer implying very skewed distributionFollowing Bisin et al (2008) code 4 (lsquoVery importantrsquo) as answer on questions9 is coded TRUE else FALSE Codes 8 (lsquoCanrsquot sayrsquo) and 9 are coded asNA (FNSEM 1993a p 112f)

gt isna(U$s9) lt- U$s9==8 | U$s9==9

gt U$ImportanceofReligion lt- U$s9 == 4

253 Attitude towards inter-marriage

Question s34a is ldquoWould you personally mind if a close relative were tomarry a white personrdquo It serves as a filter question to s34b (ldquoWould youmind very much or just a littlerdquo) which is asked to those who answered yes(cod 1) on s34a We code those who answer yes on both questions (Mind ampMind very much) as TRUE The category FALSE refers then to those who(Do not mind) or (Mind amp Mind Little) Code 8 (Canrsquot say) on s34a andcode 9 on s34a and on s34b are assigned as NA (FNSEM 1993a p 125)

gt isna(U$s34a) lt- U$s34a==8 | U$s34a==9

gt isna(U$s34b) lt- U$s34b==8

gt U$AttitudeTowardsInterMarriage lt-

+ U$s34a==1 amp U$s34b==1

7

254 Importance of racial composition in schools

Two questions are asked s23 is

ldquoIf you were choosing a school for an eleven-year old child ofyours would your choice be influenced by how many (RESPON-DENTrsquoS ETHNIC ORIGIN) children there were in the schoolrdquo(FNSEM 1993a p 120)

and s24a asks that if the available school were similar in other ways wouldyou prefer to send this child to school with fewer than half of the pupils(code 1) about half of the students (code 2) more than half (code 3) wereof your own ethnic origin s23 is not a filter question Importance of RacialComposition in Schools is set to TRUE if s24a is equal to 3 and FALSEotherwise Code 7 (No preference) is coded as FALSE Codes 8 (Canrsquot say)and 9 are assigned as NA (FNSEM 1993a p 120)

gt isna(U$s24a) lt- U$s24a==8 | U$s24a==9

gt U$ImportanceofRacialCompositioninSchools lt- U$s24a==3

255 Born in the UK

Defines who is born in the United Kingdom (question a3) Category 16 isNorthern Ireland category 17 England and Wales and category 18 ScotlandCode 99 is assigned as NA (FNSEM 1993a p 107)

gt isna(U$a3)lt- U$a3==99

gt U$BornintheUK lt- U$a3==16 | U$a3==17 | U$a3==18

256 Age at and years since arrival

This part defines the variables Age at arrival and Years since arrival by usinginformation about year of migration (question a4n) age (question a1an) andthe year (variable year when the interview is made) The interview is madein 93 or 94) The result is that some individuals get the age at arrival minus1which presumably is due to rounding of years since arrival and the age

All born in the UK are coded as 0 for AgeatArrival and YearsSince

Arrival Since these two variables are related to the age of the immigrantsone could also add an interaction variable between BornintheUK and Age

to account for effect of age for the natives The effect of age for natives is notrepresented in the model as specified by Bisin et al (2008) We experimentedwith this and results were basically unchanged The interaction variableis insignificant in all specifications Code 99 is assigned as NA for year(FNSEM 1993a pp 105 107)

8

gt isna(U$a4n) lt- U$a4n==98 | U$a4n==99

gt isna(U$year)lt- U$year==99

gt U$Age lt- U$a1an

gt U$YearsSinceArrival lt- ifelse(U$BornintheUK==TRUE

+ 0 U$year-U$a4n)

gt U$AgeatArrival lt- U$Age - U$YearsSinceArrival

gt U$AgeatArrival lt- replace(U$AgeatArrival

+ U$BornintheUK==TRUE 0)

257 Female

Definition of females via question hh2as Code na is coded as NA (FNSEM1993a pp 318)

gt isna(U$hh2as) lt- U$hh2as ==na

gt U$Female lt- U$hh2a==female

258 Arranged marriage

In question s39 Indian Pakistani Bangladeshi and Chinese respondents whohas ever been married were asked a question about the decision regardingtheir marriage The question ask about the role of the respondent and hisor her parents about choosing the respondentrsquos husbandwife In categories1 and 2 of s39 the respondentrsquos parents made the final decision and thesecategories define the dummy where the respondent is or has been living inan arranged marriage (code 1 all other are coded 0) Notice that singlesand Caribbeans have not received this question Singles are already removedfrom the sample (see section 23 Caribbeans are coded 0 This means thatthe Caribbeans do not marry according to the decision of their parentsCategory 8 (Canrsquot say) and category 9 are assigned NA (FNSEM 1993app 127)

gt isna(U$s39) lt- U$s39==8 | U$s39==9

gt U$ArrangedMarriage lt- ifelse(

+ U$s39==3 | U$s39==4 | U$s39==5 |

+ U$ethnic==caribbean FALSE TRUE)

259 Discrimination

The discrimination variable is based on a series of questions related to dis-crimination v1a-v1d about physical attacks v9a about insults j55a andj63a discrimination at work Basically anyone answering that they havebeen discriminated for any of these reasons are coded 1 else code 0

Questions v1a-v1d is a series of filter questions Question v1a asks if therespondent have been attacked (yes or no) question v1b how many attacks

9

the respondent has been enduring and question v1c asks those who havebeen attacked once if they believe the attack had to do with reasons todo with race or colour and v1d asks the same question and regards thosewho have been attacked more than once Generally code 8 and code 9are assigned NA except for question v1b where also code 7 is assigned NA(FNSEM 1993a pp 154ff 163 195 and 199)

gt vjlist lt- c(paste(v1letters[14]sep=)

+ v9aj55aj63a)

gt isna(U[vjlist]) lt- U[vjlist]==8 |

+ U[vjlist]==9

gt isna(U$v1b) lt- U$v1b==7 | U$v1b==8 |

+ U$v1b==9

gt U$Discrimination lt-

+ (U$v1a==1 amp U$v1b==1 amp U$v1c==1) |

+ (U$v1a==1 amp

+ (U$v1b gt= 2 amp U$v1b lt= 6) amp

+ U$v1d==1) | U$v9a==1 | U$j55a==1 |

+ U$j63a==1

2510 Children

No question about the number of children is asked Instead the number ofchildren has to be calculated indirectly via the number of children not livingat home (questions f16a and f16b1n-f16b3n) and the relation between therespondent and other persons living in the household (questions hh2cb-hh2cm)

The number of children out of home is calculated in the following wayIf children out of home is TRUE (f16a=1) then the number of childrenequals the sum of f16b1n f16b2n and f16b3n (the number of children notliving at home below 5 years between 5 and 15 years and above 15 yearsof age) Else if there are no children out of home (ie if f16a=2) thenthe number of children out of home is set to 0 Missing values are coded asbelow (FNSEM 1993a p 57)

gt isna(U$f16a) lt- U$f16a==8 | U$f16a==9

gt isna(U$f16b1n) lt- U$f16b1n==99

gt isna(U$f16b2n) lt- U$f16b2n==99

gt isna(U$f16b3n) lt- U$f16b3n==98 |

+ U$f16b3n==99

gt U$ChildnotatHome lt- ifelse(U$f16a==1

+ U$f16b1n+U$f16b2n+U$f16b3n0)

Questions hh2cb-hh2cm are about the relationship between the re-spondent and other individuals in the household (person b c d etc to person

10

m) category 5 being child of the respondent First we check if the personis a child to the respondent and then all children are summed over the re-spondents household adding the variable measuring number of children notat home Generally codes 98 and 99 are assigned NA (FNSEM 1993a p318)

gt hhlist lt- c(paste(hh2c letters[213] sep=))

gt isna(U[hhlist]) lt- U[hhlist]==98 | U[hhlist]==99

gt U$Children lt-

+ U$ChildnotatHome + apply(apply(subset(U

+ select=c(hh2cbhh2cm)) 2

+ function(x) x==5)1 function(x) sum(x

+ narm=TRUE))

2511 No British education

Question q1 asks whether the respondent has any British education Code2 is no Code 8 (Canacutet say) is kept in the alternative category since theseindividuals will answer the question q3 about foreign education Code 9 inq1 is assigned NA (FNSEM 1993a p 96)

gt isna(U$q1) lt- U$q1==9

gt U$NoBritishEducation lt- U$q1==2

2512 British basic education

We could not exactly see how this variable was defined in Bisin et al (2008)They define the British high education as Andashlevel and above One interpre-tation is then that O-level are educations included in the basic level Thisinterpretation is implemented here NA is assigned to all observations forwhich the filter question No British Education was NA (FNSEM 1993a pp96ff)

gt q2alist lt- c(paste(q2a c(181218) sep=))

gt U$BritishBasicEducation lt-

+ apply(apply(U[q2alist] 2function(x)

+ x==1)1 function(x) sum(x narm=TRUE))=0

gt U$BritishBasicEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishBasicEducation)

11

2513 British higher education

Bisin et al (2008) explicitly defined British higher education as A-levelGiven the definition of British Basic education the reference group willinclude trade apprenticeships as well as university educations NA is assignedto all observations for which the filter question No British Education wasNA (FNSEM 1993a pp 96ff)

gt U$BritishHigherEducation lt- apply(apply(subset(

+ Uselect=c(q2a9q2a11q2a19q2a20))

+ 2function(x) x==1)1

+ function(x) sum(x narm=TRUE))=0

gt U$BritishHigherEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishHigherEducation)

2514 Foreign education

Foreign educations is question q3 asked to all who answered lsquonorsquo or lsquoCanrsquotrsquosayrsquo on question q1 The answer yes is coded as (code 1) and no is codedas(code 2) Contrary to the above educational variables NA is not assignedto all NA on No British Education since some of them (code 8) actually wasasked the question q3 Instead NA is assigned to all observations for whichthe filter question q1 was 1 or 9 and to all code 8 (Canrsquot say) and code 9 onquestion q3 (FNSEM 1993a p 99)

gt isna(U$q3) lt- U$q3==8 | U$q3==9

gt U$ForeignEducation lt- U$q3==1

2515 Labour market status

We code the labour market status using j1b (in paid work last week or not)and j3occ (classification of activity either last weekrsquos activity or potentialactivity during the last ten years) The variable j1b takes the value 1 forpaid work last week and 2 otherwise The variable j3occ is coded as follows(FNSEM 1993ab the former pp 81f)

1 Self-employed (25+ employees)2 Self-employed (1-24 employees)3 Self-employed (no employees)4 Self-employed (employees not known)5 Manager (establishment of 25+ employees)6 Manager (establishment of 1-24 employees)7 Manager (employees not known)8 Foremansupervisor9 Other employee

12

10 Employee status unknown11 Not knownnot answered

Employee In order to be classified as an employee the individual has tohave answered yes (value 1) in j1b and be classified as employee in j3occ

(value 9) and have the value NotAssignedNA is assigned to categories 10and 11 in j3occ

gt isna(U$j3occ) lt- U$j3occ==10 | U$j3occ==11

gt U$LabourMarketStatus lt- ifelse(U$j1b==1 amp

+ U$j3occ==9 amp isna(U$j1b==1 amp

+ U$j3occ==9) EmployeeNotAssigned)

Self Employed Selfndashemployed are also coded using j1b and j3occ aboveCategories 1minus 4 in j3occ are defined as self-employed We also require thatj1b is equal 1 (SelfndashEmployed)

gt U$LabourMarketStatus lt- replace(

+ U$LabourMarketStatus

+ (U$j3occ==1 | U$j3occ==2 |

+ U$j3occ==3 | U$j3occ==4) amp

+ U$j1b==1SelfEmployed)

Manager Managers are also coded using j1b and j3occ see above Cat-egories 5 minus 8 in j3occ are defined as managers (including supervisors)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ U$j1b==1 amp U$j3occgt4 amp U$j3occlt9

+ Manager)

Unemployed The question hh5ds describes the respondentrsquos labour mar-ket status Unemployment is defined via this variable hh5ds is coded inthe following way (FNSEM 1993b)

1 Full-time education2 Govt training programme3 Full-time paid work4 Part-time paid work5 Waiting to take up paid work6 Registered unemployed7 Unemployed not registered8 Permanently sick or disabled9 Wholly retired from work

13

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

questions involved in the study are only answered by individuals who werefaced with the green questionnaire Therefore we keep only these in thesample

gt FNSEM lt- FNSEM[FNSEM$ethnic=white amp

+ isna(FNSEM$ethnic)]

gt FNSEM lt- FNSEM[FNSEM$s6=2]

gt FNSEM lt- FNSEM[FNSEM$a1e=3]

gt U lt- FNSEM lt- FNSEM[FNSEM$question==1]

An issue where Bisin et al (2008) is imprecise is whether the questionsthey address regard Muslimsnon-Muslims or Muslimnon-Muslim immi-grants Different sample selections are possible here The model specifi-cations in Bisin et al (2008) implies that White Muslims are excluded andnative ethnic minority Muslims are included This sample definition does notmatch Bisin et al (2008) writing using the terms Muslim and non-Muslimimmigrants as the sample includes natives

24 Recoding of missing values

The data set contains several codes for missing values These missing valuescan be of different characters eg nonndashavailables lsquocanrsquot sayrsquo or because therespondent was filtered in a previous filter question We employ the strategyto code all these as nonndashavailables in R ie code NA In the data set genuinenon-availables are generally coded as minus1 We set all minus1 to NA in the entiredata set

gt isna(U) lt- U==-1

In addition to minus1 several other codes (na 7 8 9 98 99 997 and 999)are occasionally used in the data set to indicate various unknown categoriesSome of minus1 and other unknown categories are nonndashavailables and have tobe deleted in estimations This is done after we have coded all our variablesin section 27 In questions following a filter question NA may have to beset in a category Some of these other codes used to denote unknowns aregenuine NArsquos and has to be removed Others will be included in a categoryThis is done variable by variable below

25 Variable definitions

Using the same variable names as in Bisin et al (2008) we define the vari-ables at the precision described by the authors We here give our interpre-tation of the variable definitions in Bisin et al (2008)

6

251 Religious affiliation

Question s6 asks whether the respondent belongs to a church or has a re-ligion For those who answer yes question s7 asks which that church orreligion is We define two religious affiliations muslim (which are all whoanswered category 3 (muslim) on question s7) and non-muslim (all who didnot answer category 3 (muslim) on question s7) All nonndashreligious that isthose who answered category 2 on question s6 are already removed from thesample Observations containing na are recoded to NA (FNSEM 1993a p112f)

gt isna(U$s7) lt- U$s7==na

gt U$Religion lt- ifelse(U$s7==muslimmuslimnon-muslim)

252 Importance of religion

Question s9 is about the importance of religion To grade the Importanceof Religion respondents have to choose between the following categories 1not at all important 2 not very important 3 fairly important or 4 veryimportant Following standard coding practice of such questions 1 and 2should be one category and 3 and 4 another but Bisin et al (2008) chooseto put 1 2 and 3 in the same category

Among those who have answered the question very few have chosenthe alternative 1 or 2 in their answer implying very skewed distributionFollowing Bisin et al (2008) code 4 (lsquoVery importantrsquo) as answer on questions9 is coded TRUE else FALSE Codes 8 (lsquoCanrsquot sayrsquo) and 9 are coded asNA (FNSEM 1993a p 112f)

gt isna(U$s9) lt- U$s9==8 | U$s9==9

gt U$ImportanceofReligion lt- U$s9 == 4

253 Attitude towards inter-marriage

Question s34a is ldquoWould you personally mind if a close relative were tomarry a white personrdquo It serves as a filter question to s34b (ldquoWould youmind very much or just a littlerdquo) which is asked to those who answered yes(cod 1) on s34a We code those who answer yes on both questions (Mind ampMind very much) as TRUE The category FALSE refers then to those who(Do not mind) or (Mind amp Mind Little) Code 8 (Canrsquot say) on s34a andcode 9 on s34a and on s34b are assigned as NA (FNSEM 1993a p 125)

gt isna(U$s34a) lt- U$s34a==8 | U$s34a==9

gt isna(U$s34b) lt- U$s34b==8

gt U$AttitudeTowardsInterMarriage lt-

+ U$s34a==1 amp U$s34b==1

7

254 Importance of racial composition in schools

Two questions are asked s23 is

ldquoIf you were choosing a school for an eleven-year old child ofyours would your choice be influenced by how many (RESPON-DENTrsquoS ETHNIC ORIGIN) children there were in the schoolrdquo(FNSEM 1993a p 120)

and s24a asks that if the available school were similar in other ways wouldyou prefer to send this child to school with fewer than half of the pupils(code 1) about half of the students (code 2) more than half (code 3) wereof your own ethnic origin s23 is not a filter question Importance of RacialComposition in Schools is set to TRUE if s24a is equal to 3 and FALSEotherwise Code 7 (No preference) is coded as FALSE Codes 8 (Canrsquot say)and 9 are assigned as NA (FNSEM 1993a p 120)

gt isna(U$s24a) lt- U$s24a==8 | U$s24a==9

gt U$ImportanceofRacialCompositioninSchools lt- U$s24a==3

255 Born in the UK

Defines who is born in the United Kingdom (question a3) Category 16 isNorthern Ireland category 17 England and Wales and category 18 ScotlandCode 99 is assigned as NA (FNSEM 1993a p 107)

gt isna(U$a3)lt- U$a3==99

gt U$BornintheUK lt- U$a3==16 | U$a3==17 | U$a3==18

256 Age at and years since arrival

This part defines the variables Age at arrival and Years since arrival by usinginformation about year of migration (question a4n) age (question a1an) andthe year (variable year when the interview is made) The interview is madein 93 or 94) The result is that some individuals get the age at arrival minus1which presumably is due to rounding of years since arrival and the age

All born in the UK are coded as 0 for AgeatArrival and YearsSince

Arrival Since these two variables are related to the age of the immigrantsone could also add an interaction variable between BornintheUK and Age

to account for effect of age for the natives The effect of age for natives is notrepresented in the model as specified by Bisin et al (2008) We experimentedwith this and results were basically unchanged The interaction variableis insignificant in all specifications Code 99 is assigned as NA for year(FNSEM 1993a pp 105 107)

8

gt isna(U$a4n) lt- U$a4n==98 | U$a4n==99

gt isna(U$year)lt- U$year==99

gt U$Age lt- U$a1an

gt U$YearsSinceArrival lt- ifelse(U$BornintheUK==TRUE

+ 0 U$year-U$a4n)

gt U$AgeatArrival lt- U$Age - U$YearsSinceArrival

gt U$AgeatArrival lt- replace(U$AgeatArrival

+ U$BornintheUK==TRUE 0)

257 Female

Definition of females via question hh2as Code na is coded as NA (FNSEM1993a pp 318)

gt isna(U$hh2as) lt- U$hh2as ==na

gt U$Female lt- U$hh2a==female

258 Arranged marriage

In question s39 Indian Pakistani Bangladeshi and Chinese respondents whohas ever been married were asked a question about the decision regardingtheir marriage The question ask about the role of the respondent and hisor her parents about choosing the respondentrsquos husbandwife In categories1 and 2 of s39 the respondentrsquos parents made the final decision and thesecategories define the dummy where the respondent is or has been living inan arranged marriage (code 1 all other are coded 0) Notice that singlesand Caribbeans have not received this question Singles are already removedfrom the sample (see section 23 Caribbeans are coded 0 This means thatthe Caribbeans do not marry according to the decision of their parentsCategory 8 (Canrsquot say) and category 9 are assigned NA (FNSEM 1993app 127)

gt isna(U$s39) lt- U$s39==8 | U$s39==9

gt U$ArrangedMarriage lt- ifelse(

+ U$s39==3 | U$s39==4 | U$s39==5 |

+ U$ethnic==caribbean FALSE TRUE)

259 Discrimination

The discrimination variable is based on a series of questions related to dis-crimination v1a-v1d about physical attacks v9a about insults j55a andj63a discrimination at work Basically anyone answering that they havebeen discriminated for any of these reasons are coded 1 else code 0

Questions v1a-v1d is a series of filter questions Question v1a asks if therespondent have been attacked (yes or no) question v1b how many attacks

9

the respondent has been enduring and question v1c asks those who havebeen attacked once if they believe the attack had to do with reasons todo with race or colour and v1d asks the same question and regards thosewho have been attacked more than once Generally code 8 and code 9are assigned NA except for question v1b where also code 7 is assigned NA(FNSEM 1993a pp 154ff 163 195 and 199)

gt vjlist lt- c(paste(v1letters[14]sep=)

+ v9aj55aj63a)

gt isna(U[vjlist]) lt- U[vjlist]==8 |

+ U[vjlist]==9

gt isna(U$v1b) lt- U$v1b==7 | U$v1b==8 |

+ U$v1b==9

gt U$Discrimination lt-

+ (U$v1a==1 amp U$v1b==1 amp U$v1c==1) |

+ (U$v1a==1 amp

+ (U$v1b gt= 2 amp U$v1b lt= 6) amp

+ U$v1d==1) | U$v9a==1 | U$j55a==1 |

+ U$j63a==1

2510 Children

No question about the number of children is asked Instead the number ofchildren has to be calculated indirectly via the number of children not livingat home (questions f16a and f16b1n-f16b3n) and the relation between therespondent and other persons living in the household (questions hh2cb-hh2cm)

The number of children out of home is calculated in the following wayIf children out of home is TRUE (f16a=1) then the number of childrenequals the sum of f16b1n f16b2n and f16b3n (the number of children notliving at home below 5 years between 5 and 15 years and above 15 yearsof age) Else if there are no children out of home (ie if f16a=2) thenthe number of children out of home is set to 0 Missing values are coded asbelow (FNSEM 1993a p 57)

gt isna(U$f16a) lt- U$f16a==8 | U$f16a==9

gt isna(U$f16b1n) lt- U$f16b1n==99

gt isna(U$f16b2n) lt- U$f16b2n==99

gt isna(U$f16b3n) lt- U$f16b3n==98 |

+ U$f16b3n==99

gt U$ChildnotatHome lt- ifelse(U$f16a==1

+ U$f16b1n+U$f16b2n+U$f16b3n0)

Questions hh2cb-hh2cm are about the relationship between the re-spondent and other individuals in the household (person b c d etc to person

10

m) category 5 being child of the respondent First we check if the personis a child to the respondent and then all children are summed over the re-spondents household adding the variable measuring number of children notat home Generally codes 98 and 99 are assigned NA (FNSEM 1993a p318)

gt hhlist lt- c(paste(hh2c letters[213] sep=))

gt isna(U[hhlist]) lt- U[hhlist]==98 | U[hhlist]==99

gt U$Children lt-

+ U$ChildnotatHome + apply(apply(subset(U

+ select=c(hh2cbhh2cm)) 2

+ function(x) x==5)1 function(x) sum(x

+ narm=TRUE))

2511 No British education

Question q1 asks whether the respondent has any British education Code2 is no Code 8 (Canacutet say) is kept in the alternative category since theseindividuals will answer the question q3 about foreign education Code 9 inq1 is assigned NA (FNSEM 1993a p 96)

gt isna(U$q1) lt- U$q1==9

gt U$NoBritishEducation lt- U$q1==2

2512 British basic education

We could not exactly see how this variable was defined in Bisin et al (2008)They define the British high education as Andashlevel and above One interpre-tation is then that O-level are educations included in the basic level Thisinterpretation is implemented here NA is assigned to all observations forwhich the filter question No British Education was NA (FNSEM 1993a pp96ff)

gt q2alist lt- c(paste(q2a c(181218) sep=))

gt U$BritishBasicEducation lt-

+ apply(apply(U[q2alist] 2function(x)

+ x==1)1 function(x) sum(x narm=TRUE))=0

gt U$BritishBasicEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishBasicEducation)

11

2513 British higher education

Bisin et al (2008) explicitly defined British higher education as A-levelGiven the definition of British Basic education the reference group willinclude trade apprenticeships as well as university educations NA is assignedto all observations for which the filter question No British Education wasNA (FNSEM 1993a pp 96ff)

gt U$BritishHigherEducation lt- apply(apply(subset(

+ Uselect=c(q2a9q2a11q2a19q2a20))

+ 2function(x) x==1)1

+ function(x) sum(x narm=TRUE))=0

gt U$BritishHigherEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishHigherEducation)

2514 Foreign education

Foreign educations is question q3 asked to all who answered lsquonorsquo or lsquoCanrsquotrsquosayrsquo on question q1 The answer yes is coded as (code 1) and no is codedas(code 2) Contrary to the above educational variables NA is not assignedto all NA on No British Education since some of them (code 8) actually wasasked the question q3 Instead NA is assigned to all observations for whichthe filter question q1 was 1 or 9 and to all code 8 (Canrsquot say) and code 9 onquestion q3 (FNSEM 1993a p 99)

gt isna(U$q3) lt- U$q3==8 | U$q3==9

gt U$ForeignEducation lt- U$q3==1

2515 Labour market status

We code the labour market status using j1b (in paid work last week or not)and j3occ (classification of activity either last weekrsquos activity or potentialactivity during the last ten years) The variable j1b takes the value 1 forpaid work last week and 2 otherwise The variable j3occ is coded as follows(FNSEM 1993ab the former pp 81f)

1 Self-employed (25+ employees)2 Self-employed (1-24 employees)3 Self-employed (no employees)4 Self-employed (employees not known)5 Manager (establishment of 25+ employees)6 Manager (establishment of 1-24 employees)7 Manager (employees not known)8 Foremansupervisor9 Other employee

12

10 Employee status unknown11 Not knownnot answered

Employee In order to be classified as an employee the individual has tohave answered yes (value 1) in j1b and be classified as employee in j3occ

(value 9) and have the value NotAssignedNA is assigned to categories 10and 11 in j3occ

gt isna(U$j3occ) lt- U$j3occ==10 | U$j3occ==11

gt U$LabourMarketStatus lt- ifelse(U$j1b==1 amp

+ U$j3occ==9 amp isna(U$j1b==1 amp

+ U$j3occ==9) EmployeeNotAssigned)

Self Employed Selfndashemployed are also coded using j1b and j3occ aboveCategories 1minus 4 in j3occ are defined as self-employed We also require thatj1b is equal 1 (SelfndashEmployed)

gt U$LabourMarketStatus lt- replace(

+ U$LabourMarketStatus

+ (U$j3occ==1 | U$j3occ==2 |

+ U$j3occ==3 | U$j3occ==4) amp

+ U$j1b==1SelfEmployed)

Manager Managers are also coded using j1b and j3occ see above Cat-egories 5 minus 8 in j3occ are defined as managers (including supervisors)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ U$j1b==1 amp U$j3occgt4 amp U$j3occlt9

+ Manager)

Unemployed The question hh5ds describes the respondentrsquos labour mar-ket status Unemployment is defined via this variable hh5ds is coded inthe following way (FNSEM 1993b)

1 Full-time education2 Govt training programme3 Full-time paid work4 Part-time paid work5 Waiting to take up paid work6 Registered unemployed7 Unemployed not registered8 Permanently sick or disabled9 Wholly retired from work

13

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

251 Religious affiliation

Question s6 asks whether the respondent belongs to a church or has a re-ligion For those who answer yes question s7 asks which that church orreligion is We define two religious affiliations muslim (which are all whoanswered category 3 (muslim) on question s7) and non-muslim (all who didnot answer category 3 (muslim) on question s7) All nonndashreligious that isthose who answered category 2 on question s6 are already removed from thesample Observations containing na are recoded to NA (FNSEM 1993a p112f)

gt isna(U$s7) lt- U$s7==na

gt U$Religion lt- ifelse(U$s7==muslimmuslimnon-muslim)

252 Importance of religion

Question s9 is about the importance of religion To grade the Importanceof Religion respondents have to choose between the following categories 1not at all important 2 not very important 3 fairly important or 4 veryimportant Following standard coding practice of such questions 1 and 2should be one category and 3 and 4 another but Bisin et al (2008) chooseto put 1 2 and 3 in the same category

Among those who have answered the question very few have chosenthe alternative 1 or 2 in their answer implying very skewed distributionFollowing Bisin et al (2008) code 4 (lsquoVery importantrsquo) as answer on questions9 is coded TRUE else FALSE Codes 8 (lsquoCanrsquot sayrsquo) and 9 are coded asNA (FNSEM 1993a p 112f)

gt isna(U$s9) lt- U$s9==8 | U$s9==9

gt U$ImportanceofReligion lt- U$s9 == 4

253 Attitude towards inter-marriage

Question s34a is ldquoWould you personally mind if a close relative were tomarry a white personrdquo It serves as a filter question to s34b (ldquoWould youmind very much or just a littlerdquo) which is asked to those who answered yes(cod 1) on s34a We code those who answer yes on both questions (Mind ampMind very much) as TRUE The category FALSE refers then to those who(Do not mind) or (Mind amp Mind Little) Code 8 (Canrsquot say) on s34a andcode 9 on s34a and on s34b are assigned as NA (FNSEM 1993a p 125)

gt isna(U$s34a) lt- U$s34a==8 | U$s34a==9

gt isna(U$s34b) lt- U$s34b==8

gt U$AttitudeTowardsInterMarriage lt-

+ U$s34a==1 amp U$s34b==1

7

254 Importance of racial composition in schools

Two questions are asked s23 is

ldquoIf you were choosing a school for an eleven-year old child ofyours would your choice be influenced by how many (RESPON-DENTrsquoS ETHNIC ORIGIN) children there were in the schoolrdquo(FNSEM 1993a p 120)

and s24a asks that if the available school were similar in other ways wouldyou prefer to send this child to school with fewer than half of the pupils(code 1) about half of the students (code 2) more than half (code 3) wereof your own ethnic origin s23 is not a filter question Importance of RacialComposition in Schools is set to TRUE if s24a is equal to 3 and FALSEotherwise Code 7 (No preference) is coded as FALSE Codes 8 (Canrsquot say)and 9 are assigned as NA (FNSEM 1993a p 120)

gt isna(U$s24a) lt- U$s24a==8 | U$s24a==9

gt U$ImportanceofRacialCompositioninSchools lt- U$s24a==3

255 Born in the UK

Defines who is born in the United Kingdom (question a3) Category 16 isNorthern Ireland category 17 England and Wales and category 18 ScotlandCode 99 is assigned as NA (FNSEM 1993a p 107)

gt isna(U$a3)lt- U$a3==99

gt U$BornintheUK lt- U$a3==16 | U$a3==17 | U$a3==18

256 Age at and years since arrival

This part defines the variables Age at arrival and Years since arrival by usinginformation about year of migration (question a4n) age (question a1an) andthe year (variable year when the interview is made) The interview is madein 93 or 94) The result is that some individuals get the age at arrival minus1which presumably is due to rounding of years since arrival and the age

All born in the UK are coded as 0 for AgeatArrival and YearsSince

Arrival Since these two variables are related to the age of the immigrantsone could also add an interaction variable between BornintheUK and Age

to account for effect of age for the natives The effect of age for natives is notrepresented in the model as specified by Bisin et al (2008) We experimentedwith this and results were basically unchanged The interaction variableis insignificant in all specifications Code 99 is assigned as NA for year(FNSEM 1993a pp 105 107)

8

gt isna(U$a4n) lt- U$a4n==98 | U$a4n==99

gt isna(U$year)lt- U$year==99

gt U$Age lt- U$a1an

gt U$YearsSinceArrival lt- ifelse(U$BornintheUK==TRUE

+ 0 U$year-U$a4n)

gt U$AgeatArrival lt- U$Age - U$YearsSinceArrival

gt U$AgeatArrival lt- replace(U$AgeatArrival

+ U$BornintheUK==TRUE 0)

257 Female

Definition of females via question hh2as Code na is coded as NA (FNSEM1993a pp 318)

gt isna(U$hh2as) lt- U$hh2as ==na

gt U$Female lt- U$hh2a==female

258 Arranged marriage

In question s39 Indian Pakistani Bangladeshi and Chinese respondents whohas ever been married were asked a question about the decision regardingtheir marriage The question ask about the role of the respondent and hisor her parents about choosing the respondentrsquos husbandwife In categories1 and 2 of s39 the respondentrsquos parents made the final decision and thesecategories define the dummy where the respondent is or has been living inan arranged marriage (code 1 all other are coded 0) Notice that singlesand Caribbeans have not received this question Singles are already removedfrom the sample (see section 23 Caribbeans are coded 0 This means thatthe Caribbeans do not marry according to the decision of their parentsCategory 8 (Canrsquot say) and category 9 are assigned NA (FNSEM 1993app 127)

gt isna(U$s39) lt- U$s39==8 | U$s39==9

gt U$ArrangedMarriage lt- ifelse(

+ U$s39==3 | U$s39==4 | U$s39==5 |

+ U$ethnic==caribbean FALSE TRUE)

259 Discrimination

The discrimination variable is based on a series of questions related to dis-crimination v1a-v1d about physical attacks v9a about insults j55a andj63a discrimination at work Basically anyone answering that they havebeen discriminated for any of these reasons are coded 1 else code 0

Questions v1a-v1d is a series of filter questions Question v1a asks if therespondent have been attacked (yes or no) question v1b how many attacks

9

the respondent has been enduring and question v1c asks those who havebeen attacked once if they believe the attack had to do with reasons todo with race or colour and v1d asks the same question and regards thosewho have been attacked more than once Generally code 8 and code 9are assigned NA except for question v1b where also code 7 is assigned NA(FNSEM 1993a pp 154ff 163 195 and 199)

gt vjlist lt- c(paste(v1letters[14]sep=)

+ v9aj55aj63a)

gt isna(U[vjlist]) lt- U[vjlist]==8 |

+ U[vjlist]==9

gt isna(U$v1b) lt- U$v1b==7 | U$v1b==8 |

+ U$v1b==9

gt U$Discrimination lt-

+ (U$v1a==1 amp U$v1b==1 amp U$v1c==1) |

+ (U$v1a==1 amp

+ (U$v1b gt= 2 amp U$v1b lt= 6) amp

+ U$v1d==1) | U$v9a==1 | U$j55a==1 |

+ U$j63a==1

2510 Children

No question about the number of children is asked Instead the number ofchildren has to be calculated indirectly via the number of children not livingat home (questions f16a and f16b1n-f16b3n) and the relation between therespondent and other persons living in the household (questions hh2cb-hh2cm)

The number of children out of home is calculated in the following wayIf children out of home is TRUE (f16a=1) then the number of childrenequals the sum of f16b1n f16b2n and f16b3n (the number of children notliving at home below 5 years between 5 and 15 years and above 15 yearsof age) Else if there are no children out of home (ie if f16a=2) thenthe number of children out of home is set to 0 Missing values are coded asbelow (FNSEM 1993a p 57)

gt isna(U$f16a) lt- U$f16a==8 | U$f16a==9

gt isna(U$f16b1n) lt- U$f16b1n==99

gt isna(U$f16b2n) lt- U$f16b2n==99

gt isna(U$f16b3n) lt- U$f16b3n==98 |

+ U$f16b3n==99

gt U$ChildnotatHome lt- ifelse(U$f16a==1

+ U$f16b1n+U$f16b2n+U$f16b3n0)

Questions hh2cb-hh2cm are about the relationship between the re-spondent and other individuals in the household (person b c d etc to person

10

m) category 5 being child of the respondent First we check if the personis a child to the respondent and then all children are summed over the re-spondents household adding the variable measuring number of children notat home Generally codes 98 and 99 are assigned NA (FNSEM 1993a p318)

gt hhlist lt- c(paste(hh2c letters[213] sep=))

gt isna(U[hhlist]) lt- U[hhlist]==98 | U[hhlist]==99

gt U$Children lt-

+ U$ChildnotatHome + apply(apply(subset(U

+ select=c(hh2cbhh2cm)) 2

+ function(x) x==5)1 function(x) sum(x

+ narm=TRUE))

2511 No British education

Question q1 asks whether the respondent has any British education Code2 is no Code 8 (Canacutet say) is kept in the alternative category since theseindividuals will answer the question q3 about foreign education Code 9 inq1 is assigned NA (FNSEM 1993a p 96)

gt isna(U$q1) lt- U$q1==9

gt U$NoBritishEducation lt- U$q1==2

2512 British basic education

We could not exactly see how this variable was defined in Bisin et al (2008)They define the British high education as Andashlevel and above One interpre-tation is then that O-level are educations included in the basic level Thisinterpretation is implemented here NA is assigned to all observations forwhich the filter question No British Education was NA (FNSEM 1993a pp96ff)

gt q2alist lt- c(paste(q2a c(181218) sep=))

gt U$BritishBasicEducation lt-

+ apply(apply(U[q2alist] 2function(x)

+ x==1)1 function(x) sum(x narm=TRUE))=0

gt U$BritishBasicEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishBasicEducation)

11

2513 British higher education

Bisin et al (2008) explicitly defined British higher education as A-levelGiven the definition of British Basic education the reference group willinclude trade apprenticeships as well as university educations NA is assignedto all observations for which the filter question No British Education wasNA (FNSEM 1993a pp 96ff)

gt U$BritishHigherEducation lt- apply(apply(subset(

+ Uselect=c(q2a9q2a11q2a19q2a20))

+ 2function(x) x==1)1

+ function(x) sum(x narm=TRUE))=0

gt U$BritishHigherEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishHigherEducation)

2514 Foreign education

Foreign educations is question q3 asked to all who answered lsquonorsquo or lsquoCanrsquotrsquosayrsquo on question q1 The answer yes is coded as (code 1) and no is codedas(code 2) Contrary to the above educational variables NA is not assignedto all NA on No British Education since some of them (code 8) actually wasasked the question q3 Instead NA is assigned to all observations for whichthe filter question q1 was 1 or 9 and to all code 8 (Canrsquot say) and code 9 onquestion q3 (FNSEM 1993a p 99)

gt isna(U$q3) lt- U$q3==8 | U$q3==9

gt U$ForeignEducation lt- U$q3==1

2515 Labour market status

We code the labour market status using j1b (in paid work last week or not)and j3occ (classification of activity either last weekrsquos activity or potentialactivity during the last ten years) The variable j1b takes the value 1 forpaid work last week and 2 otherwise The variable j3occ is coded as follows(FNSEM 1993ab the former pp 81f)

1 Self-employed (25+ employees)2 Self-employed (1-24 employees)3 Self-employed (no employees)4 Self-employed (employees not known)5 Manager (establishment of 25+ employees)6 Manager (establishment of 1-24 employees)7 Manager (employees not known)8 Foremansupervisor9 Other employee

12

10 Employee status unknown11 Not knownnot answered

Employee In order to be classified as an employee the individual has tohave answered yes (value 1) in j1b and be classified as employee in j3occ

(value 9) and have the value NotAssignedNA is assigned to categories 10and 11 in j3occ

gt isna(U$j3occ) lt- U$j3occ==10 | U$j3occ==11

gt U$LabourMarketStatus lt- ifelse(U$j1b==1 amp

+ U$j3occ==9 amp isna(U$j1b==1 amp

+ U$j3occ==9) EmployeeNotAssigned)

Self Employed Selfndashemployed are also coded using j1b and j3occ aboveCategories 1minus 4 in j3occ are defined as self-employed We also require thatj1b is equal 1 (SelfndashEmployed)

gt U$LabourMarketStatus lt- replace(

+ U$LabourMarketStatus

+ (U$j3occ==1 | U$j3occ==2 |

+ U$j3occ==3 | U$j3occ==4) amp

+ U$j1b==1SelfEmployed)

Manager Managers are also coded using j1b and j3occ see above Cat-egories 5 minus 8 in j3occ are defined as managers (including supervisors)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ U$j1b==1 amp U$j3occgt4 amp U$j3occlt9

+ Manager)

Unemployed The question hh5ds describes the respondentrsquos labour mar-ket status Unemployment is defined via this variable hh5ds is coded inthe following way (FNSEM 1993b)

1 Full-time education2 Govt training programme3 Full-time paid work4 Part-time paid work5 Waiting to take up paid work6 Registered unemployed7 Unemployed not registered8 Permanently sick or disabled9 Wholly retired from work

13

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

254 Importance of racial composition in schools

Two questions are asked s23 is

ldquoIf you were choosing a school for an eleven-year old child ofyours would your choice be influenced by how many (RESPON-DENTrsquoS ETHNIC ORIGIN) children there were in the schoolrdquo(FNSEM 1993a p 120)

and s24a asks that if the available school were similar in other ways wouldyou prefer to send this child to school with fewer than half of the pupils(code 1) about half of the students (code 2) more than half (code 3) wereof your own ethnic origin s23 is not a filter question Importance of RacialComposition in Schools is set to TRUE if s24a is equal to 3 and FALSEotherwise Code 7 (No preference) is coded as FALSE Codes 8 (Canrsquot say)and 9 are assigned as NA (FNSEM 1993a p 120)

gt isna(U$s24a) lt- U$s24a==8 | U$s24a==9

gt U$ImportanceofRacialCompositioninSchools lt- U$s24a==3

255 Born in the UK

Defines who is born in the United Kingdom (question a3) Category 16 isNorthern Ireland category 17 England and Wales and category 18 ScotlandCode 99 is assigned as NA (FNSEM 1993a p 107)

gt isna(U$a3)lt- U$a3==99

gt U$BornintheUK lt- U$a3==16 | U$a3==17 | U$a3==18

256 Age at and years since arrival

This part defines the variables Age at arrival and Years since arrival by usinginformation about year of migration (question a4n) age (question a1an) andthe year (variable year when the interview is made) The interview is madein 93 or 94) The result is that some individuals get the age at arrival minus1which presumably is due to rounding of years since arrival and the age

All born in the UK are coded as 0 for AgeatArrival and YearsSince

Arrival Since these two variables are related to the age of the immigrantsone could also add an interaction variable between BornintheUK and Age

to account for effect of age for the natives The effect of age for natives is notrepresented in the model as specified by Bisin et al (2008) We experimentedwith this and results were basically unchanged The interaction variableis insignificant in all specifications Code 99 is assigned as NA for year(FNSEM 1993a pp 105 107)

8

gt isna(U$a4n) lt- U$a4n==98 | U$a4n==99

gt isna(U$year)lt- U$year==99

gt U$Age lt- U$a1an

gt U$YearsSinceArrival lt- ifelse(U$BornintheUK==TRUE

+ 0 U$year-U$a4n)

gt U$AgeatArrival lt- U$Age - U$YearsSinceArrival

gt U$AgeatArrival lt- replace(U$AgeatArrival

+ U$BornintheUK==TRUE 0)

257 Female

Definition of females via question hh2as Code na is coded as NA (FNSEM1993a pp 318)

gt isna(U$hh2as) lt- U$hh2as ==na

gt U$Female lt- U$hh2a==female

258 Arranged marriage

In question s39 Indian Pakistani Bangladeshi and Chinese respondents whohas ever been married were asked a question about the decision regardingtheir marriage The question ask about the role of the respondent and hisor her parents about choosing the respondentrsquos husbandwife In categories1 and 2 of s39 the respondentrsquos parents made the final decision and thesecategories define the dummy where the respondent is or has been living inan arranged marriage (code 1 all other are coded 0) Notice that singlesand Caribbeans have not received this question Singles are already removedfrom the sample (see section 23 Caribbeans are coded 0 This means thatthe Caribbeans do not marry according to the decision of their parentsCategory 8 (Canrsquot say) and category 9 are assigned NA (FNSEM 1993app 127)

gt isna(U$s39) lt- U$s39==8 | U$s39==9

gt U$ArrangedMarriage lt- ifelse(

+ U$s39==3 | U$s39==4 | U$s39==5 |

+ U$ethnic==caribbean FALSE TRUE)

259 Discrimination

The discrimination variable is based on a series of questions related to dis-crimination v1a-v1d about physical attacks v9a about insults j55a andj63a discrimination at work Basically anyone answering that they havebeen discriminated for any of these reasons are coded 1 else code 0

Questions v1a-v1d is a series of filter questions Question v1a asks if therespondent have been attacked (yes or no) question v1b how many attacks

9

the respondent has been enduring and question v1c asks those who havebeen attacked once if they believe the attack had to do with reasons todo with race or colour and v1d asks the same question and regards thosewho have been attacked more than once Generally code 8 and code 9are assigned NA except for question v1b where also code 7 is assigned NA(FNSEM 1993a pp 154ff 163 195 and 199)

gt vjlist lt- c(paste(v1letters[14]sep=)

+ v9aj55aj63a)

gt isna(U[vjlist]) lt- U[vjlist]==8 |

+ U[vjlist]==9

gt isna(U$v1b) lt- U$v1b==7 | U$v1b==8 |

+ U$v1b==9

gt U$Discrimination lt-

+ (U$v1a==1 amp U$v1b==1 amp U$v1c==1) |

+ (U$v1a==1 amp

+ (U$v1b gt= 2 amp U$v1b lt= 6) amp

+ U$v1d==1) | U$v9a==1 | U$j55a==1 |

+ U$j63a==1

2510 Children

No question about the number of children is asked Instead the number ofchildren has to be calculated indirectly via the number of children not livingat home (questions f16a and f16b1n-f16b3n) and the relation between therespondent and other persons living in the household (questions hh2cb-hh2cm)

The number of children out of home is calculated in the following wayIf children out of home is TRUE (f16a=1) then the number of childrenequals the sum of f16b1n f16b2n and f16b3n (the number of children notliving at home below 5 years between 5 and 15 years and above 15 yearsof age) Else if there are no children out of home (ie if f16a=2) thenthe number of children out of home is set to 0 Missing values are coded asbelow (FNSEM 1993a p 57)

gt isna(U$f16a) lt- U$f16a==8 | U$f16a==9

gt isna(U$f16b1n) lt- U$f16b1n==99

gt isna(U$f16b2n) lt- U$f16b2n==99

gt isna(U$f16b3n) lt- U$f16b3n==98 |

+ U$f16b3n==99

gt U$ChildnotatHome lt- ifelse(U$f16a==1

+ U$f16b1n+U$f16b2n+U$f16b3n0)

Questions hh2cb-hh2cm are about the relationship between the re-spondent and other individuals in the household (person b c d etc to person

10

m) category 5 being child of the respondent First we check if the personis a child to the respondent and then all children are summed over the re-spondents household adding the variable measuring number of children notat home Generally codes 98 and 99 are assigned NA (FNSEM 1993a p318)

gt hhlist lt- c(paste(hh2c letters[213] sep=))

gt isna(U[hhlist]) lt- U[hhlist]==98 | U[hhlist]==99

gt U$Children lt-

+ U$ChildnotatHome + apply(apply(subset(U

+ select=c(hh2cbhh2cm)) 2

+ function(x) x==5)1 function(x) sum(x

+ narm=TRUE))

2511 No British education

Question q1 asks whether the respondent has any British education Code2 is no Code 8 (Canacutet say) is kept in the alternative category since theseindividuals will answer the question q3 about foreign education Code 9 inq1 is assigned NA (FNSEM 1993a p 96)

gt isna(U$q1) lt- U$q1==9

gt U$NoBritishEducation lt- U$q1==2

2512 British basic education

We could not exactly see how this variable was defined in Bisin et al (2008)They define the British high education as Andashlevel and above One interpre-tation is then that O-level are educations included in the basic level Thisinterpretation is implemented here NA is assigned to all observations forwhich the filter question No British Education was NA (FNSEM 1993a pp96ff)

gt q2alist lt- c(paste(q2a c(181218) sep=))

gt U$BritishBasicEducation lt-

+ apply(apply(U[q2alist] 2function(x)

+ x==1)1 function(x) sum(x narm=TRUE))=0

gt U$BritishBasicEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishBasicEducation)

11

2513 British higher education

Bisin et al (2008) explicitly defined British higher education as A-levelGiven the definition of British Basic education the reference group willinclude trade apprenticeships as well as university educations NA is assignedto all observations for which the filter question No British Education wasNA (FNSEM 1993a pp 96ff)

gt U$BritishHigherEducation lt- apply(apply(subset(

+ Uselect=c(q2a9q2a11q2a19q2a20))

+ 2function(x) x==1)1

+ function(x) sum(x narm=TRUE))=0

gt U$BritishHigherEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishHigherEducation)

2514 Foreign education

Foreign educations is question q3 asked to all who answered lsquonorsquo or lsquoCanrsquotrsquosayrsquo on question q1 The answer yes is coded as (code 1) and no is codedas(code 2) Contrary to the above educational variables NA is not assignedto all NA on No British Education since some of them (code 8) actually wasasked the question q3 Instead NA is assigned to all observations for whichthe filter question q1 was 1 or 9 and to all code 8 (Canrsquot say) and code 9 onquestion q3 (FNSEM 1993a p 99)

gt isna(U$q3) lt- U$q3==8 | U$q3==9

gt U$ForeignEducation lt- U$q3==1

2515 Labour market status

We code the labour market status using j1b (in paid work last week or not)and j3occ (classification of activity either last weekrsquos activity or potentialactivity during the last ten years) The variable j1b takes the value 1 forpaid work last week and 2 otherwise The variable j3occ is coded as follows(FNSEM 1993ab the former pp 81f)

1 Self-employed (25+ employees)2 Self-employed (1-24 employees)3 Self-employed (no employees)4 Self-employed (employees not known)5 Manager (establishment of 25+ employees)6 Manager (establishment of 1-24 employees)7 Manager (employees not known)8 Foremansupervisor9 Other employee

12

10 Employee status unknown11 Not knownnot answered

Employee In order to be classified as an employee the individual has tohave answered yes (value 1) in j1b and be classified as employee in j3occ

(value 9) and have the value NotAssignedNA is assigned to categories 10and 11 in j3occ

gt isna(U$j3occ) lt- U$j3occ==10 | U$j3occ==11

gt U$LabourMarketStatus lt- ifelse(U$j1b==1 amp

+ U$j3occ==9 amp isna(U$j1b==1 amp

+ U$j3occ==9) EmployeeNotAssigned)

Self Employed Selfndashemployed are also coded using j1b and j3occ aboveCategories 1minus 4 in j3occ are defined as self-employed We also require thatj1b is equal 1 (SelfndashEmployed)

gt U$LabourMarketStatus lt- replace(

+ U$LabourMarketStatus

+ (U$j3occ==1 | U$j3occ==2 |

+ U$j3occ==3 | U$j3occ==4) amp

+ U$j1b==1SelfEmployed)

Manager Managers are also coded using j1b and j3occ see above Cat-egories 5 minus 8 in j3occ are defined as managers (including supervisors)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ U$j1b==1 amp U$j3occgt4 amp U$j3occlt9

+ Manager)

Unemployed The question hh5ds describes the respondentrsquos labour mar-ket status Unemployment is defined via this variable hh5ds is coded inthe following way (FNSEM 1993b)

1 Full-time education2 Govt training programme3 Full-time paid work4 Part-time paid work5 Waiting to take up paid work6 Registered unemployed7 Unemployed not registered8 Permanently sick or disabled9 Wholly retired from work

13

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

gt isna(U$a4n) lt- U$a4n==98 | U$a4n==99

gt isna(U$year)lt- U$year==99

gt U$Age lt- U$a1an

gt U$YearsSinceArrival lt- ifelse(U$BornintheUK==TRUE

+ 0 U$year-U$a4n)

gt U$AgeatArrival lt- U$Age - U$YearsSinceArrival

gt U$AgeatArrival lt- replace(U$AgeatArrival

+ U$BornintheUK==TRUE 0)

257 Female

Definition of females via question hh2as Code na is coded as NA (FNSEM1993a pp 318)

gt isna(U$hh2as) lt- U$hh2as ==na

gt U$Female lt- U$hh2a==female

258 Arranged marriage

In question s39 Indian Pakistani Bangladeshi and Chinese respondents whohas ever been married were asked a question about the decision regardingtheir marriage The question ask about the role of the respondent and hisor her parents about choosing the respondentrsquos husbandwife In categories1 and 2 of s39 the respondentrsquos parents made the final decision and thesecategories define the dummy where the respondent is or has been living inan arranged marriage (code 1 all other are coded 0) Notice that singlesand Caribbeans have not received this question Singles are already removedfrom the sample (see section 23 Caribbeans are coded 0 This means thatthe Caribbeans do not marry according to the decision of their parentsCategory 8 (Canrsquot say) and category 9 are assigned NA (FNSEM 1993app 127)

gt isna(U$s39) lt- U$s39==8 | U$s39==9

gt U$ArrangedMarriage lt- ifelse(

+ U$s39==3 | U$s39==4 | U$s39==5 |

+ U$ethnic==caribbean FALSE TRUE)

259 Discrimination

The discrimination variable is based on a series of questions related to dis-crimination v1a-v1d about physical attacks v9a about insults j55a andj63a discrimination at work Basically anyone answering that they havebeen discriminated for any of these reasons are coded 1 else code 0

Questions v1a-v1d is a series of filter questions Question v1a asks if therespondent have been attacked (yes or no) question v1b how many attacks

9

the respondent has been enduring and question v1c asks those who havebeen attacked once if they believe the attack had to do with reasons todo with race or colour and v1d asks the same question and regards thosewho have been attacked more than once Generally code 8 and code 9are assigned NA except for question v1b where also code 7 is assigned NA(FNSEM 1993a pp 154ff 163 195 and 199)

gt vjlist lt- c(paste(v1letters[14]sep=)

+ v9aj55aj63a)

gt isna(U[vjlist]) lt- U[vjlist]==8 |

+ U[vjlist]==9

gt isna(U$v1b) lt- U$v1b==7 | U$v1b==8 |

+ U$v1b==9

gt U$Discrimination lt-

+ (U$v1a==1 amp U$v1b==1 amp U$v1c==1) |

+ (U$v1a==1 amp

+ (U$v1b gt= 2 amp U$v1b lt= 6) amp

+ U$v1d==1) | U$v9a==1 | U$j55a==1 |

+ U$j63a==1

2510 Children

No question about the number of children is asked Instead the number ofchildren has to be calculated indirectly via the number of children not livingat home (questions f16a and f16b1n-f16b3n) and the relation between therespondent and other persons living in the household (questions hh2cb-hh2cm)

The number of children out of home is calculated in the following wayIf children out of home is TRUE (f16a=1) then the number of childrenequals the sum of f16b1n f16b2n and f16b3n (the number of children notliving at home below 5 years between 5 and 15 years and above 15 yearsof age) Else if there are no children out of home (ie if f16a=2) thenthe number of children out of home is set to 0 Missing values are coded asbelow (FNSEM 1993a p 57)

gt isna(U$f16a) lt- U$f16a==8 | U$f16a==9

gt isna(U$f16b1n) lt- U$f16b1n==99

gt isna(U$f16b2n) lt- U$f16b2n==99

gt isna(U$f16b3n) lt- U$f16b3n==98 |

+ U$f16b3n==99

gt U$ChildnotatHome lt- ifelse(U$f16a==1

+ U$f16b1n+U$f16b2n+U$f16b3n0)

Questions hh2cb-hh2cm are about the relationship between the re-spondent and other individuals in the household (person b c d etc to person

10

m) category 5 being child of the respondent First we check if the personis a child to the respondent and then all children are summed over the re-spondents household adding the variable measuring number of children notat home Generally codes 98 and 99 are assigned NA (FNSEM 1993a p318)

gt hhlist lt- c(paste(hh2c letters[213] sep=))

gt isna(U[hhlist]) lt- U[hhlist]==98 | U[hhlist]==99

gt U$Children lt-

+ U$ChildnotatHome + apply(apply(subset(U

+ select=c(hh2cbhh2cm)) 2

+ function(x) x==5)1 function(x) sum(x

+ narm=TRUE))

2511 No British education

Question q1 asks whether the respondent has any British education Code2 is no Code 8 (Canacutet say) is kept in the alternative category since theseindividuals will answer the question q3 about foreign education Code 9 inq1 is assigned NA (FNSEM 1993a p 96)

gt isna(U$q1) lt- U$q1==9

gt U$NoBritishEducation lt- U$q1==2

2512 British basic education

We could not exactly see how this variable was defined in Bisin et al (2008)They define the British high education as Andashlevel and above One interpre-tation is then that O-level are educations included in the basic level Thisinterpretation is implemented here NA is assigned to all observations forwhich the filter question No British Education was NA (FNSEM 1993a pp96ff)

gt q2alist lt- c(paste(q2a c(181218) sep=))

gt U$BritishBasicEducation lt-

+ apply(apply(U[q2alist] 2function(x)

+ x==1)1 function(x) sum(x narm=TRUE))=0

gt U$BritishBasicEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishBasicEducation)

11

2513 British higher education

Bisin et al (2008) explicitly defined British higher education as A-levelGiven the definition of British Basic education the reference group willinclude trade apprenticeships as well as university educations NA is assignedto all observations for which the filter question No British Education wasNA (FNSEM 1993a pp 96ff)

gt U$BritishHigherEducation lt- apply(apply(subset(

+ Uselect=c(q2a9q2a11q2a19q2a20))

+ 2function(x) x==1)1

+ function(x) sum(x narm=TRUE))=0

gt U$BritishHigherEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishHigherEducation)

2514 Foreign education

Foreign educations is question q3 asked to all who answered lsquonorsquo or lsquoCanrsquotrsquosayrsquo on question q1 The answer yes is coded as (code 1) and no is codedas(code 2) Contrary to the above educational variables NA is not assignedto all NA on No British Education since some of them (code 8) actually wasasked the question q3 Instead NA is assigned to all observations for whichthe filter question q1 was 1 or 9 and to all code 8 (Canrsquot say) and code 9 onquestion q3 (FNSEM 1993a p 99)

gt isna(U$q3) lt- U$q3==8 | U$q3==9

gt U$ForeignEducation lt- U$q3==1

2515 Labour market status

We code the labour market status using j1b (in paid work last week or not)and j3occ (classification of activity either last weekrsquos activity or potentialactivity during the last ten years) The variable j1b takes the value 1 forpaid work last week and 2 otherwise The variable j3occ is coded as follows(FNSEM 1993ab the former pp 81f)

1 Self-employed (25+ employees)2 Self-employed (1-24 employees)3 Self-employed (no employees)4 Self-employed (employees not known)5 Manager (establishment of 25+ employees)6 Manager (establishment of 1-24 employees)7 Manager (employees not known)8 Foremansupervisor9 Other employee

12

10 Employee status unknown11 Not knownnot answered

Employee In order to be classified as an employee the individual has tohave answered yes (value 1) in j1b and be classified as employee in j3occ

(value 9) and have the value NotAssignedNA is assigned to categories 10and 11 in j3occ

gt isna(U$j3occ) lt- U$j3occ==10 | U$j3occ==11

gt U$LabourMarketStatus lt- ifelse(U$j1b==1 amp

+ U$j3occ==9 amp isna(U$j1b==1 amp

+ U$j3occ==9) EmployeeNotAssigned)

Self Employed Selfndashemployed are also coded using j1b and j3occ aboveCategories 1minus 4 in j3occ are defined as self-employed We also require thatj1b is equal 1 (SelfndashEmployed)

gt U$LabourMarketStatus lt- replace(

+ U$LabourMarketStatus

+ (U$j3occ==1 | U$j3occ==2 |

+ U$j3occ==3 | U$j3occ==4) amp

+ U$j1b==1SelfEmployed)

Manager Managers are also coded using j1b and j3occ see above Cat-egories 5 minus 8 in j3occ are defined as managers (including supervisors)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ U$j1b==1 amp U$j3occgt4 amp U$j3occlt9

+ Manager)

Unemployed The question hh5ds describes the respondentrsquos labour mar-ket status Unemployment is defined via this variable hh5ds is coded inthe following way (FNSEM 1993b)

1 Full-time education2 Govt training programme3 Full-time paid work4 Part-time paid work5 Waiting to take up paid work6 Registered unemployed7 Unemployed not registered8 Permanently sick or disabled9 Wholly retired from work

13

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

the respondent has been enduring and question v1c asks those who havebeen attacked once if they believe the attack had to do with reasons todo with race or colour and v1d asks the same question and regards thosewho have been attacked more than once Generally code 8 and code 9are assigned NA except for question v1b where also code 7 is assigned NA(FNSEM 1993a pp 154ff 163 195 and 199)

gt vjlist lt- c(paste(v1letters[14]sep=)

+ v9aj55aj63a)

gt isna(U[vjlist]) lt- U[vjlist]==8 |

+ U[vjlist]==9

gt isna(U$v1b) lt- U$v1b==7 | U$v1b==8 |

+ U$v1b==9

gt U$Discrimination lt-

+ (U$v1a==1 amp U$v1b==1 amp U$v1c==1) |

+ (U$v1a==1 amp

+ (U$v1b gt= 2 amp U$v1b lt= 6) amp

+ U$v1d==1) | U$v9a==1 | U$j55a==1 |

+ U$j63a==1

2510 Children

No question about the number of children is asked Instead the number ofchildren has to be calculated indirectly via the number of children not livingat home (questions f16a and f16b1n-f16b3n) and the relation between therespondent and other persons living in the household (questions hh2cb-hh2cm)

The number of children out of home is calculated in the following wayIf children out of home is TRUE (f16a=1) then the number of childrenequals the sum of f16b1n f16b2n and f16b3n (the number of children notliving at home below 5 years between 5 and 15 years and above 15 yearsof age) Else if there are no children out of home (ie if f16a=2) thenthe number of children out of home is set to 0 Missing values are coded asbelow (FNSEM 1993a p 57)

gt isna(U$f16a) lt- U$f16a==8 | U$f16a==9

gt isna(U$f16b1n) lt- U$f16b1n==99

gt isna(U$f16b2n) lt- U$f16b2n==99

gt isna(U$f16b3n) lt- U$f16b3n==98 |

+ U$f16b3n==99

gt U$ChildnotatHome lt- ifelse(U$f16a==1

+ U$f16b1n+U$f16b2n+U$f16b3n0)

Questions hh2cb-hh2cm are about the relationship between the re-spondent and other individuals in the household (person b c d etc to person

10

m) category 5 being child of the respondent First we check if the personis a child to the respondent and then all children are summed over the re-spondents household adding the variable measuring number of children notat home Generally codes 98 and 99 are assigned NA (FNSEM 1993a p318)

gt hhlist lt- c(paste(hh2c letters[213] sep=))

gt isna(U[hhlist]) lt- U[hhlist]==98 | U[hhlist]==99

gt U$Children lt-

+ U$ChildnotatHome + apply(apply(subset(U

+ select=c(hh2cbhh2cm)) 2

+ function(x) x==5)1 function(x) sum(x

+ narm=TRUE))

2511 No British education

Question q1 asks whether the respondent has any British education Code2 is no Code 8 (Canacutet say) is kept in the alternative category since theseindividuals will answer the question q3 about foreign education Code 9 inq1 is assigned NA (FNSEM 1993a p 96)

gt isna(U$q1) lt- U$q1==9

gt U$NoBritishEducation lt- U$q1==2

2512 British basic education

We could not exactly see how this variable was defined in Bisin et al (2008)They define the British high education as Andashlevel and above One interpre-tation is then that O-level are educations included in the basic level Thisinterpretation is implemented here NA is assigned to all observations forwhich the filter question No British Education was NA (FNSEM 1993a pp96ff)

gt q2alist lt- c(paste(q2a c(181218) sep=))

gt U$BritishBasicEducation lt-

+ apply(apply(U[q2alist] 2function(x)

+ x==1)1 function(x) sum(x narm=TRUE))=0

gt U$BritishBasicEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishBasicEducation)

11

2513 British higher education

Bisin et al (2008) explicitly defined British higher education as A-levelGiven the definition of British Basic education the reference group willinclude trade apprenticeships as well as university educations NA is assignedto all observations for which the filter question No British Education wasNA (FNSEM 1993a pp 96ff)

gt U$BritishHigherEducation lt- apply(apply(subset(

+ Uselect=c(q2a9q2a11q2a19q2a20))

+ 2function(x) x==1)1

+ function(x) sum(x narm=TRUE))=0

gt U$BritishHigherEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishHigherEducation)

2514 Foreign education

Foreign educations is question q3 asked to all who answered lsquonorsquo or lsquoCanrsquotrsquosayrsquo on question q1 The answer yes is coded as (code 1) and no is codedas(code 2) Contrary to the above educational variables NA is not assignedto all NA on No British Education since some of them (code 8) actually wasasked the question q3 Instead NA is assigned to all observations for whichthe filter question q1 was 1 or 9 and to all code 8 (Canrsquot say) and code 9 onquestion q3 (FNSEM 1993a p 99)

gt isna(U$q3) lt- U$q3==8 | U$q3==9

gt U$ForeignEducation lt- U$q3==1

2515 Labour market status

We code the labour market status using j1b (in paid work last week or not)and j3occ (classification of activity either last weekrsquos activity or potentialactivity during the last ten years) The variable j1b takes the value 1 forpaid work last week and 2 otherwise The variable j3occ is coded as follows(FNSEM 1993ab the former pp 81f)

1 Self-employed (25+ employees)2 Self-employed (1-24 employees)3 Self-employed (no employees)4 Self-employed (employees not known)5 Manager (establishment of 25+ employees)6 Manager (establishment of 1-24 employees)7 Manager (employees not known)8 Foremansupervisor9 Other employee

12

10 Employee status unknown11 Not knownnot answered

Employee In order to be classified as an employee the individual has tohave answered yes (value 1) in j1b and be classified as employee in j3occ

(value 9) and have the value NotAssignedNA is assigned to categories 10and 11 in j3occ

gt isna(U$j3occ) lt- U$j3occ==10 | U$j3occ==11

gt U$LabourMarketStatus lt- ifelse(U$j1b==1 amp

+ U$j3occ==9 amp isna(U$j1b==1 amp

+ U$j3occ==9) EmployeeNotAssigned)

Self Employed Selfndashemployed are also coded using j1b and j3occ aboveCategories 1minus 4 in j3occ are defined as self-employed We also require thatj1b is equal 1 (SelfndashEmployed)

gt U$LabourMarketStatus lt- replace(

+ U$LabourMarketStatus

+ (U$j3occ==1 | U$j3occ==2 |

+ U$j3occ==3 | U$j3occ==4) amp

+ U$j1b==1SelfEmployed)

Manager Managers are also coded using j1b and j3occ see above Cat-egories 5 minus 8 in j3occ are defined as managers (including supervisors)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ U$j1b==1 amp U$j3occgt4 amp U$j3occlt9

+ Manager)

Unemployed The question hh5ds describes the respondentrsquos labour mar-ket status Unemployment is defined via this variable hh5ds is coded inthe following way (FNSEM 1993b)

1 Full-time education2 Govt training programme3 Full-time paid work4 Part-time paid work5 Waiting to take up paid work6 Registered unemployed7 Unemployed not registered8 Permanently sick or disabled9 Wholly retired from work

13

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

m) category 5 being child of the respondent First we check if the personis a child to the respondent and then all children are summed over the re-spondents household adding the variable measuring number of children notat home Generally codes 98 and 99 are assigned NA (FNSEM 1993a p318)

gt hhlist lt- c(paste(hh2c letters[213] sep=))

gt isna(U[hhlist]) lt- U[hhlist]==98 | U[hhlist]==99

gt U$Children lt-

+ U$ChildnotatHome + apply(apply(subset(U

+ select=c(hh2cbhh2cm)) 2

+ function(x) x==5)1 function(x) sum(x

+ narm=TRUE))

2511 No British education

Question q1 asks whether the respondent has any British education Code2 is no Code 8 (Canacutet say) is kept in the alternative category since theseindividuals will answer the question q3 about foreign education Code 9 inq1 is assigned NA (FNSEM 1993a p 96)

gt isna(U$q1) lt- U$q1==9

gt U$NoBritishEducation lt- U$q1==2

2512 British basic education

We could not exactly see how this variable was defined in Bisin et al (2008)They define the British high education as Andashlevel and above One interpre-tation is then that O-level are educations included in the basic level Thisinterpretation is implemented here NA is assigned to all observations forwhich the filter question No British Education was NA (FNSEM 1993a pp96ff)

gt q2alist lt- c(paste(q2a c(181218) sep=))

gt U$BritishBasicEducation lt-

+ apply(apply(U[q2alist] 2function(x)

+ x==1)1 function(x) sum(x narm=TRUE))=0

gt U$BritishBasicEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishBasicEducation)

11

2513 British higher education

Bisin et al (2008) explicitly defined British higher education as A-levelGiven the definition of British Basic education the reference group willinclude trade apprenticeships as well as university educations NA is assignedto all observations for which the filter question No British Education wasNA (FNSEM 1993a pp 96ff)

gt U$BritishHigherEducation lt- apply(apply(subset(

+ Uselect=c(q2a9q2a11q2a19q2a20))

+ 2function(x) x==1)1

+ function(x) sum(x narm=TRUE))=0

gt U$BritishHigherEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishHigherEducation)

2514 Foreign education

Foreign educations is question q3 asked to all who answered lsquonorsquo or lsquoCanrsquotrsquosayrsquo on question q1 The answer yes is coded as (code 1) and no is codedas(code 2) Contrary to the above educational variables NA is not assignedto all NA on No British Education since some of them (code 8) actually wasasked the question q3 Instead NA is assigned to all observations for whichthe filter question q1 was 1 or 9 and to all code 8 (Canrsquot say) and code 9 onquestion q3 (FNSEM 1993a p 99)

gt isna(U$q3) lt- U$q3==8 | U$q3==9

gt U$ForeignEducation lt- U$q3==1

2515 Labour market status

We code the labour market status using j1b (in paid work last week or not)and j3occ (classification of activity either last weekrsquos activity or potentialactivity during the last ten years) The variable j1b takes the value 1 forpaid work last week and 2 otherwise The variable j3occ is coded as follows(FNSEM 1993ab the former pp 81f)

1 Self-employed (25+ employees)2 Self-employed (1-24 employees)3 Self-employed (no employees)4 Self-employed (employees not known)5 Manager (establishment of 25+ employees)6 Manager (establishment of 1-24 employees)7 Manager (employees not known)8 Foremansupervisor9 Other employee

12

10 Employee status unknown11 Not knownnot answered

Employee In order to be classified as an employee the individual has tohave answered yes (value 1) in j1b and be classified as employee in j3occ

(value 9) and have the value NotAssignedNA is assigned to categories 10and 11 in j3occ

gt isna(U$j3occ) lt- U$j3occ==10 | U$j3occ==11

gt U$LabourMarketStatus lt- ifelse(U$j1b==1 amp

+ U$j3occ==9 amp isna(U$j1b==1 amp

+ U$j3occ==9) EmployeeNotAssigned)

Self Employed Selfndashemployed are also coded using j1b and j3occ aboveCategories 1minus 4 in j3occ are defined as self-employed We also require thatj1b is equal 1 (SelfndashEmployed)

gt U$LabourMarketStatus lt- replace(

+ U$LabourMarketStatus

+ (U$j3occ==1 | U$j3occ==2 |

+ U$j3occ==3 | U$j3occ==4) amp

+ U$j1b==1SelfEmployed)

Manager Managers are also coded using j1b and j3occ see above Cat-egories 5 minus 8 in j3occ are defined as managers (including supervisors)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ U$j1b==1 amp U$j3occgt4 amp U$j3occlt9

+ Manager)

Unemployed The question hh5ds describes the respondentrsquos labour mar-ket status Unemployment is defined via this variable hh5ds is coded inthe following way (FNSEM 1993b)

1 Full-time education2 Govt training programme3 Full-time paid work4 Part-time paid work5 Waiting to take up paid work6 Registered unemployed7 Unemployed not registered8 Permanently sick or disabled9 Wholly retired from work

13

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

2513 British higher education

Bisin et al (2008) explicitly defined British higher education as A-levelGiven the definition of British Basic education the reference group willinclude trade apprenticeships as well as university educations NA is assignedto all observations for which the filter question No British Education wasNA (FNSEM 1993a pp 96ff)

gt U$BritishHigherEducation lt- apply(apply(subset(

+ Uselect=c(q2a9q2a11q2a19q2a20))

+ 2function(x) x==1)1

+ function(x) sum(x narm=TRUE))=0

gt U$BritishHigherEducation lt- ifelse(

+ isna(U$NoBritishEducation)

+ NAU$BritishHigherEducation)

2514 Foreign education

Foreign educations is question q3 asked to all who answered lsquonorsquo or lsquoCanrsquotrsquosayrsquo on question q1 The answer yes is coded as (code 1) and no is codedas(code 2) Contrary to the above educational variables NA is not assignedto all NA on No British Education since some of them (code 8) actually wasasked the question q3 Instead NA is assigned to all observations for whichthe filter question q1 was 1 or 9 and to all code 8 (Canrsquot say) and code 9 onquestion q3 (FNSEM 1993a p 99)

gt isna(U$q3) lt- U$q3==8 | U$q3==9

gt U$ForeignEducation lt- U$q3==1

2515 Labour market status

We code the labour market status using j1b (in paid work last week or not)and j3occ (classification of activity either last weekrsquos activity or potentialactivity during the last ten years) The variable j1b takes the value 1 forpaid work last week and 2 otherwise The variable j3occ is coded as follows(FNSEM 1993ab the former pp 81f)

1 Self-employed (25+ employees)2 Self-employed (1-24 employees)3 Self-employed (no employees)4 Self-employed (employees not known)5 Manager (establishment of 25+ employees)6 Manager (establishment of 1-24 employees)7 Manager (employees not known)8 Foremansupervisor9 Other employee

12

10 Employee status unknown11 Not knownnot answered

Employee In order to be classified as an employee the individual has tohave answered yes (value 1) in j1b and be classified as employee in j3occ

(value 9) and have the value NotAssignedNA is assigned to categories 10and 11 in j3occ

gt isna(U$j3occ) lt- U$j3occ==10 | U$j3occ==11

gt U$LabourMarketStatus lt- ifelse(U$j1b==1 amp

+ U$j3occ==9 amp isna(U$j1b==1 amp

+ U$j3occ==9) EmployeeNotAssigned)

Self Employed Selfndashemployed are also coded using j1b and j3occ aboveCategories 1minus 4 in j3occ are defined as self-employed We also require thatj1b is equal 1 (SelfndashEmployed)

gt U$LabourMarketStatus lt- replace(

+ U$LabourMarketStatus

+ (U$j3occ==1 | U$j3occ==2 |

+ U$j3occ==3 | U$j3occ==4) amp

+ U$j1b==1SelfEmployed)

Manager Managers are also coded using j1b and j3occ see above Cat-egories 5 minus 8 in j3occ are defined as managers (including supervisors)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ U$j1b==1 amp U$j3occgt4 amp U$j3occlt9

+ Manager)

Unemployed The question hh5ds describes the respondentrsquos labour mar-ket status Unemployment is defined via this variable hh5ds is coded inthe following way (FNSEM 1993b)

1 Full-time education2 Govt training programme3 Full-time paid work4 Part-time paid work5 Waiting to take up paid work6 Registered unemployed7 Unemployed not registered8 Permanently sick or disabled9 Wholly retired from work

13

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

10 Employee status unknown11 Not knownnot answered

Employee In order to be classified as an employee the individual has tohave answered yes (value 1) in j1b and be classified as employee in j3occ

(value 9) and have the value NotAssignedNA is assigned to categories 10and 11 in j3occ

gt isna(U$j3occ) lt- U$j3occ==10 | U$j3occ==11

gt U$LabourMarketStatus lt- ifelse(U$j1b==1 amp

+ U$j3occ==9 amp isna(U$j1b==1 amp

+ U$j3occ==9) EmployeeNotAssigned)

Self Employed Selfndashemployed are also coded using j1b and j3occ aboveCategories 1minus 4 in j3occ are defined as self-employed We also require thatj1b is equal 1 (SelfndashEmployed)

gt U$LabourMarketStatus lt- replace(

+ U$LabourMarketStatus

+ (U$j3occ==1 | U$j3occ==2 |

+ U$j3occ==3 | U$j3occ==4) amp

+ U$j1b==1SelfEmployed)

Manager Managers are also coded using j1b and j3occ see above Cat-egories 5 minus 8 in j3occ are defined as managers (including supervisors)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ U$j1b==1 amp U$j3occgt4 amp U$j3occlt9

+ Manager)

Unemployed The question hh5ds describes the respondentrsquos labour mar-ket status Unemployment is defined via this variable hh5ds is coded inthe following way (FNSEM 1993b)

1 Full-time education2 Govt training programme3 Full-time paid work4 Part-time paid work5 Waiting to take up paid work6 Registered unemployed7 Unemployed not registered8 Permanently sick or disabled9 Wholly retired from work

13

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

10 Looking after the home11 Doing something else12 NA

We define unemployed as category 6 and 7 (Bisin et al 2008 p 324)There are few cases where the individual is classified as Employee accordingto our definition above and is reported to be unemployed in hh5ds Thiscan for example be part-time unemployment We classify these individualsas having Unclear labour market status These will be checked later andbe removed if they are few in the final sample

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus=NotAssigned

+ Unclear)

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==6 | U$hh5ds == 7) amp

+ U$LabourMarketStatus==NotAssigned

+ Unemployed)

Out of labour force The category out of labour force is defined as thosehaving values (125891011) in hh5ds or value 2 in j1b or j2

gt U$LabourMarketStatus lt-

+ replace(U$LabourMarketStatus

+ (U$hh5ds==1 | U$hh5ds == 2 |

+ U$hh5ds==5 | U$hh5ds == 8 |

+ U$hh5ds==9 | U$hh5ds == 10|

+ U$hh5ds==11 | U$j1b == 2 |

+ U$j2 == 2 ) amp

+ U$LabourMarketStatus==NotAssigned

+ OutOfLabourForce)

Remaining observations with the value NotAssigned in Labour Mar-

ketStatus will be assigned NA At this point there are 4 observations codedas Unclear These are now recoded as NA

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == NotAssigned

gt isna(U$LabourMarketStatus) lt-

+ U$LabourMarketStatus == Unclear

We will create dummy variables using this variable before we run ourmodels

14

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

2516 No parents

Variable No parents means that the respondent is not living with his orher parents This variable is coded with question f14a (which takes code 1for both alive code 2 father alive code 3 for mother alive and 8 for bothdead) and f14b (which takes code 2 if both living parents do not live withthe respondent and code 6 if the only living parent does not live with therespondent else it takes one of the values 1 or 3 minus 5) No parents shouldbe coded TRUE if either both parents are dead or the respondent does notlive with any living parents However since we follow Bisin et al (2008) wecode this variable including only those who have both their parents dead orboth paprents live away from the respondent This definition implies thatthose who have a parent living away and one parent dead are assigned thevalue FALSE Code 8 and 9 are assigned NA (FNSEM 1993a p 56)

gt isna(U$f14a) lt- U$f14a==9

gt isna(U$f14b) lt- U$f14b==9

gt U$NoParents lt- U$f14a==8 | U$f14b==2

2517 Contacts with parents

The three variables measuring contacts with parents are defined via threequestions asking about the number of physical contacts (question f15bn)the number of contacts via telephone (question f15cn) and the number ofcontacts via letters (question f15dn) that the respondent has had with hisor her parents during the last four weeks conditional on not both parentsbeing dead All three takes the value of the underlaying variable if at leastone parent is alive and the value 0 if both parents are dead Code 999 onall three variables and code 997 on f15bn are assigned NA (FNSEM 1993ap 56)

gt isna(U$f15bn) lt- U$f15bn==997 | U$f15bn==999

gt isna(U$f15cn) lt- U$f15cn==999

gt isna(U$f15dn) lt- U$f15dn==999

gt U$ParentsPhysicalContacts lt- ifelse(

+ U$f14a=8 U$f15bn0)

gt U$ParentsTelephoneCalls lt- ifelse(

+ U$f14a=8 U$f15cn0)

gt U$ParentsLetters lt- ifelse(

+ U$f14a=8 U$f15dn0)

2518 English language

There are several language variables measuring whether the respondent isspeaking English with different individuals at home with older at home

15

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

with younger at work and with friends We construct theses variables usingquestion s12a which asks whether the respondent regularly speak to anyonein Britain in any other language than English and s12f s12g s12hand s12i which asks which language is spoken to the above mentionedcategories of individuals Each of s12f s12g s12h and s12i comesin 18 versions (eg s12f1 s12f18) where each question 1ndash15 iscoded yes if the respondent speaks the language Question 16 is ldquoNeverspeaks to these peopleNot ap[plicable]rdquo question 17 NA and question 18ldquo None of the above answered positiverdquo Question 1 is always regardingEnglish

The respondent is coded as English speaker if either s12a is answerednegatively or s12aX1 where X=fghi is answered positively Onlyrespondents for which either s12a is NA or all of s12X1 s12X16 areanswered negatively are coded as NA Below is the code for English SpokenAt Home With Older

gt isna(U$s12a) lt- U$s12a== 8 | U$s12a == 9

gt s12flist lt- c(paste(s12f 216 sep=))

gt U$oOLD lt- apply(U[s12flist]==yes 1 sum)

gt U$EnglishSpokenatHomewithOlder lt-

+ ((U$s12a==1 | isna(U$s12a)) amp

+ U$s12f1==yes) | U$s12a==2

gt isna(U$EnglishSpokenatHomewithOlder) lt-

+ U$oOLD==0 amp

+ U$EnglishSpokenatHomewithOlder==FALSE

gt U$EnglishSpokenatHomewithOlder lt-

+ replace(U$EnglishSpokenatHomewithOlder

+ U$oOLDgt0 amp

+ isna(U$EnglishSpokenatHomewithOlder)FALSE)

gt U$DONOTSPEAKWITHOLDER lt- ifelse(U$s12f16==no01)

The codes for English Spoken At Home With Younger English SpokenAt Work and English Spoken With Friends are equivalent These codesare in araietal_sourceRnw but not shown here It is also available inaraietal_sourceR

2519 Household income

The question hh40 provides information in which interval the householdincome of the respondentrsquos household is We assign the midpoints in theseintervals as the household income For the lowest bracket this income is themidpoint of [0 77] For the highest bracket we assign the income which is thelowest income in the bracket plus the income interval down to the midpintof the second highest bracket ie 789 + 788minus731

2= 8175 This method

16

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

underestimates the range of the highest bracket but to a lesser extent thanthe lower limit 789

gt isna(U$hh40) lt- U$hh40==refused |

+ U$hh40==cant say | U$hh40==na

gt U$HouseholdIncome lt- c(3859651350

+ 173523052600318036604140

+ 471552955870649570207595

+ 8175)[U$hh40]

2520 Ward variables

Ward density of own ethnic group is measured by the variable wown in theoriginal data This variable is coded 1 minus 7 depending on the density of therespondentrsquos own ethnic group is in the ward of the respondent (FNSEM1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-24996 25-32997 33 or more

We recode the variable to take the midpoints of the density intervals in thesame fashion as the household income variable was recoded This meansthat if the respondent is the lowest interval the density is set to be 1 etc Inthe highest interval we set the density to be the lowest density in the intervalplus the density distance down to the midpoint of the second highest densityinterval ie 33 + 33minus25

2= 37

The variable wunemp is coded 1minus 6 depending on unemployment rate inthe ward of the respondent (FNSEM 1993b)

1 Up to 1992 2-4993 5-9994 10-14995 15-206 20 or more

We recode this variable to instead take the midpoints of the intervals aswe did for the household income This means that if the respondent is thelowest interval the the rate is set to be 1 etc In the highest interval we setthe rate to be the lowest rate in the interval plus the distance down to themidpoint of the second highest rate interval ie 20 + 20minus15

2= 225

17

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

gt U$WardDensityOwnEthnicity lt- (c(1035

+ 75125200290370)[U$wown])100

gt U$WardUnemploymentRate lt- c(103575125

+ 175225)[U$wunemp]

26 Discrimination own ethnicity

Finally we define a variable describing the discrimination against the ownethnic group It is defined as the average of the variable Discrimination overethnic groups after the removal of nonndashavailables

gt GroupDiscrimination lt-

+ tapply(U$DiscriminationU$ethnic

+ function(x) mean(x narm=TRUE))

gt U$DiscriminationOwnEthnicity lt-

+ GroupDiscrimination[U$ethnic]

27 Defining the subset

We define dummy variables for labour market status to have the same vari-able labels as in Bisin et al (2008)

gt U$Employee lt- asnumeric(U$LabourMarketStatus==Employee)

gt U$Manager lt- asnumeric(U$LabourMarketStatus==Manager)

gt U$SelfEmployed lt-

+ asnumeric(U$LabourMarketStatus==SelfEmployed)

gt U$OUTOFLABOURFORCE lt-

+ asnumeric(U$LabourMarketStatus==OutOfLabourForce)

gt U$Unemployed lt-

+ asnumeric(U$LabourMarketStatus==Unemployed)

We save a data set keeping all observations containing nonndashavailables

gt UOriginal lt- U

We then choose the variables to keep in U This code is in araietal_sourceRnw

but not shown here It is also available in araietal_sourceRFinally we remove all observations containing nonndashavailables from U

gt U lt- naomit(U)

28 Sample statistics

Table 1 compares the number of observations in this sample before (seesection 27 ie data frame UOriginal) and after (see section 27 iedata frame U) nonndashavailables are removed with the numbers of observations

18

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

reported by the Bisin et al (2008) study The number of observations forvarious groups in the nonndashMuslim category are not reported in the Bisinet al (2008) paper These numbers are therefore missing in the table Thecategory definitions are from the original dataset and involves no recodingon our part

After removing observations with missing values on all variables of inter-est (ldquoAfterrdquo in Table 1) we are left with 330 Muslims and 488 non-MuslimsThe sample selections induced by the choice of variables and the missingvalues in these variables lead to a loss of 57 percent of the relevant sampleof the original data3

The sample means reported in Bisin et al (2008) seem to be unweightedSince data instructions says that the data should always be weighted tables2 and 3 report weighted and unweighted sample means before and afterremoval of non-availables and the Bisin et al (2008) data4 Comparingmeans the Bisin et al (2008) data seem to be different from the originalsample The variables Attitude Towards Inter Marriage and Importance ofRacial Composition in Schools in Bisin et al (2008) data deviate largelyfrom corresponding averages in the original data The deviation is extremein case of Importance of Racial Composition in Schools The original samplehas a mean for this variable that is 2 percent for non-Muslims (compare with33 percent in Bisin et al (2008)) and 6 percent for Muslims (compare with65 percent in Bisin et al (2008)) Due to this extremely skewed distributionit is hardly meaningfull to run a regression on this variable

Notice that also the distribution of the variable Importance of Religionwould be extremely skewed using standard coding of this type of variablesSuch a coding would imply that religion is important when the respondentanswer ldquoVery Importantrdquo and ldquoFairly Importantrdquo to the question ldquoHow im-portant is religion to the way you live your liferdquo

The sample means in our data after removing accumulated missing val-ues due to all variables in the estimations deviate marginally in general fromthe original data The similarities here are partly due to the fact that thestatistics are based exactly on the same variable definition in our implemen-tation In some respects the deviations are larger For further comparisonswe refer to Tables 2 and 3

Due to the fact that the large majority of observations from the originaldata are lost the remaining sample is likely to be contaminated with sampleselection bias To compare the characteristics of the remaining sample withthe original sample says something about systematic attrition with respectto observables The sample selection bias with respect to unobservablescannot however be resolved

3The variables written in capital letters are created to ensure wellndashdefined referencecategories They are included in our regressions but we cannot say whether they areincluded in the regressions of Bisin et al (2008)

4See FNSEM (1993b)

19

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

3 Regression Results

We use linear probability models (LPM)5 Our results are presented in Tables4 and 5 Bisin et al (2008) write that

1 ldquoMuslims integrate less and more slowly than non-Muslimsrdquo (abstractp 445) and

2 ldquo there is no evidence that segregated neighbourhoods breed intensereligious and cultural identities On the contrary intense identitiesin our data are more prominent in relatively mixed neighbourhoodsrdquo(p 446)

The first claim is based on their reported results concerning the variableYears Since Arrival In this way Bisin et al (2008) compare cohorts ofMuslims and non Muslims and attempt to say something about the evolutionof values over time They do not follow individuals over time but nonethelesscall these cohort differences ldquoIntegration over timerdquo They report negativecoefficients for Years Since Arrival but the estimates are smaller in absolutevalue for Muslims than for non-Muslims In our case the coefficients forYears Since Arrival reported in Tables 4 and 5 are insignificant in all casesexcept in the regression for Importance of Religion for Muslims where it isnegative This is opposite to what Bisin et al (2008) claim

The second claim is based on their reported results concerning the vari-able Ward Density Own Ethnicity Bisin et al (2008) report negative andsignificant estimates for Ward Density Own Ethnicity in all six specifica-tions Their negative coefficient for this variable would imply that ethnicminorities put more weight on religion mind more about inter-ethnic mar-riage and have stronger taste for ethnically profiled schools as we move fromneighbourhoods (Wards) with high density of their own ethnicity to neigh-bourhoods where people from their own ethnicity are scarce This is not atall what we find in our replication

5Bisin et al (2008) use probit estimations Our attempts to use probit run into con-vergence problems The convergence problems are severe for the model using Importance

of Racial Composition in Schools as dependent variable Hence our choice of LPMAnother issue is that Bisin et al (2008) should have included dummy variables indicating

religious affiliation Christians Sikhs and others in the non-Muslim category to checksimilarities and differences among non-Muslims as well In this respect we follow theirmodel specification

Moreover Bisin et al (2008) should have adjusted for within ward correlations Thismight matter for their standard errors which might be underestimated In our case withalmost no significant results this would not matter much The variable is not available inthe data set and we did not make much effort to obtain it

All estimated models include 7 UK-region dummies and the variables DONOT

SPEAKWITHOLDER DONOTSPEAKATWORK and DONOTSPEAKWITHFRIENDS It turnedout that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and can-not be included in the model

20

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

In our estimations the estimated coefficients for this variable are allpositive but far from significant The P-values are 069 097 and 059 forMuslims and 015 014 and 03 for non-Muslims contradicting the Bisin et al(2008) results

Inspecting the results presented in Tables 4 and 5 there are many sim-ilarities and few differences in the estimated coefficients for Muslims andnon-Muslims Our results are generally very different from results reportedby Bisin et al (2008) We are however doubtful whether it is possibleto draw any reliable inference from these results due to great loss of ob-servations and possible sample selection bias together with the problem ofendogeneity (also mentioned by Bisin et al (2008))

4 Concluding remarks

The Bisin et al (2008) paper rests on fragile grounds Our examinationof the data using their variable definitions and the same set-up indicatesthat their claims about differences between Muslims and non-Muslims andtheir conclusion that strong ReligiousEthnic identities are found in mixedneighbourhoods does not hold There is no systematic relation betweenethnic minoritiesrsquo views on religion inter-ethnic marriage or ethnic profile ofschools and the density of their own ethnic minority in their neighbourhoodHowever we hesitate to draw inference from these results since the great lossof observations (57 percent) implies that the remaining sample is most likelynot representative

5 Production notes

To facilitate reproducibility and save others timely interpretations of whatis done in this paper we attempt to follow Literate Statistical Programmingprocedures For documenting our results we have used Sweave by Leisch(2002) in combination with the LATEX family of programs using the packagesinputenc fontenc natbib Sweave fancyvrb color url hyperref andmultirow

All code (R code LATEX code and BibTEX data base code) used to do theeconometric estimations to produce this technical documentation includingall tables and to produce the companion paper Arai et al (2008) is containedin the file araietal_sourceRnw

The estimations and the documentation can be reproduced in by follow-ing the instructions at the top of the file httppeoplesuse~lundh

fragile_groundsaraietal_sourceRnw

Our results were obtained on a i486-pc-linux-gnu platform using R ver-sion 281 beta (2008-12-12 r47173) (R Development Core Team 2008) with

21

packages lmtest 09-22 (2008-12-09) sandwich 21-0 (2008-01-26) zoo 15-4 (2008-07-09) foreign 08-29 (2008-08-07) and xtable 15-4 (20081003)

References

Mahmood Arai Jonas Karlsson and Michael Lundholm On fragile groundsA replication of Are Muslim immigrants different in terms of cultural in-tegration Unpublished manuscript 2008

RG Berthoud T Modood P Smith and G Prior Fourth National Sur-vey of Ethnic Minorities 1993-1994 [computer file] Colchester EssexUK 1997 URL httpwwwdata-archiveacukdoc36855Cmrdoc

5CUKDA5CUKDA_Study_3685_Informationhtm SN 3685

Alberto Bisin Eleonora Patacchini Thierry Verdier and Yves Zenou Aremuslim immigrants different in terms of cultural integration Journal ofthe European Economic Association 6445ndash456 2008

FNSEM P1312 Fourth National Study of Etnic Minorities Projectinstructions Social and Community Planning Research London1993a URL httpwwwdata-archiveacukdoc3685mrdocpdf

a3685uabpdf Project 3685

FNSEM UK Data Archive Data Dictionary Social andCommunity Plan-ning Research London 1993b An RTF file called UKDA-3685-tab

mrdocallissue3685_UKDA_Data_Dictionaryrtf available when theFNSEM data set is downloaded from the UK Data Archive

Roger Koenker and Achim Zeileis Reproducible econometric research (acritical review of the state of the art) Report 60 Department of Statisticsand Mathematics Wirtschaftsuniversitat Wien Research Report Series2007 URL httpepubwu-wienacatdynvirlibwpengmediate

epub-wu-01_c75pdfID=epub-wu-01_c75

Friedrich Leisch Sweave user manual 2002 URL httpwwwcituwien

acat~leischSweave

R Development Core Team R A Language and Environment for StatisticalComputing R Foundation for Statistical Computing Vienna Austria2008 URL httpwwwR-projectorg ISBN 3-900051-07-0

22

Appendix Tables

Table 1 Religious affiliation (absolute () and relative () numbers) before(columns 1 and 2) and after (columns 3 and 4) removal of NA compared withBisin et al (2008) (columns 5 and 6)

Religious affiliation (1) (2) (3) (4) (5) (6)

Before After Bisin et aln = 1901 n = 818 n = 5963

hindu 359 1888 149 1822sikh 288 1515 86 1051muslim 852 4482 330 4034 2369 3973christian 357 1878 232 2836buddhist 17 089 9 110confucian 1 005 1 012jain 7 037 3 037parsizorastrian 3 016 2 024rastafarian 2 011 1 012jewish 1 005 0 000other 10 053 5 061

na 0 000 0 000NArsquos 4 021 0 000

All non-Muslims 1045 5509 488 5966 3594 6027

NOTE The row names shows exactly how the original data is coded sothat eg lsquoNArsquosrsquo are true missing values whereas lsquonarsquo is coded as religiousaffiliation lsquonarsquo On the last line nonndashMuslims are calculated excluding naand NA

23

Table 2 Weighted and Unweighted Means for Muslims and nonndashMuslims before and after removal of NA compared with Bisinet al (2008)

Muslim Non-Muslim

W e i g h t e d U n w e i g h t e d W e i g h t e d U n w e i g h t e dBefore After Before After Bisin Before After Before After Bisin

et al et al

Importance of Religion 079 081 080 085 079 049 048 048 048 042Attitude Towards Inter Marriage 045 039 045 039 070 019 015 021 017 037Importance of Racial Composition in Schools 006 008 006 007 065 002 001 002 002 033Age at Arrival 2249 2073 2236 2032 3918 2231 1785 2283 1871 4257Age 4042 4201 4018 4056 4440 4398 4490 4459Female 050 047 050 045 047 052 053 052 052 048Born in the UK 008 008 006 006 021 013 018 012 016 028Arranged Marriage 065 063 065 063 022 027 018 030 022 012Discrimination 023 020 020 016 017 038 041 035 039 019Children 312 341 323 343 217 237 232 244 236 168Years Since Arrival 1978 1927 1899 1879 2643 2419 2068 2406 2095 2708No British Education 080 079 083 082 081 056 050 059 053 052British Basic Education 014 013 012 011 006 030 036 029 035 013British Higher Education 005 006 003 003 008 010 012 009 012 016Foreign Education 025 025 023 022 025 026 024 028 025 029Employee 019 021 019 022 038 040 042 040 042 059Manager 002 002 002 002 002 007 008 006 007 004

24

Table 3 Table 2 continued Weighted and Unweighted Means for Muslims and nonndashMuslims before and after removal of NAcompared with Bisin et al (2008)

Muslim Non-Muslim

W e i g h t e d U n w e i g h t e d W e i g h t e d U n w e i g h t e dBefore After Before After Bisin Before After Before After Bisin

et al et al

Self Employed 009 008 007 006 009 012 012 010 010 014OUT OF LABOUR FORCE 050 049 052 048 033 029 036 032Unemployed 019 019 020 022 019 008 010 008 009 008No Parents 065 070 067 070 034 064 065 066 066 032Parents Physical Contacts 269 272 280 288 305 275 302 283 297 387Parents Telephone Calls 210 251 228 285 338 334 334 315 326 474Parents Letters 047 041 048 042 067 025 024 025 027 037English Spoken at Home with Older 012 012 010 009 003 043 053 042 052 008DO NOT SPEAK WITH OLDER 002 004 001 002 002 001 002 002English Spoken at Home with Younger 046 046 041 040 020 077 081 075 079 025DO NOT SPEAK WITH YOUNGER 000 000 000 000 000 000 000 000English Spoken at Work 046 053 041 047 019 079 085 077 082 027DO NOT SPEAK AT WORK 045 039 048 044 016 011 018 014English Spoken With Friends 050 055 046 051 022 079 086 077 083 027DO NOT SPEAK WITH FRIENDS 012 007 012 006 004 002 005 002Household Income 21619 21548 19540 19494 20074 33466 33383 30769 31346 33026Ward Density Own Ethnicity 013 013 016 016 015 011 010 012 011 011Ward Unemployment Rate 1607 1572 1797 1805 1657 1219 1235 1344 1352 1260Discrimination Own Ethnicity 023 022 021 019 021 034 037 034 036 018

25

Table 4 Regression Results for Muslims 330 and non-Muslims 488 to be compared with Table 2 in Bisin et al (2008)Heteroskedasticity corrected (HC1) Standard Errors are in parentheses P-values lt 005 are marked with

Importance of Inter Ethnic Ethnic CompositionReligion Marriage of Schools

Muslims nonndashMuslims Muslims nonndashMuslims Muslims nonndashMuslims

(Intercept) 092 094 083 027 004 -002(030) (026) (037) (026) (018) (004)

Age at Arrival 000 001 000 000 000 000(000) (000) (000) (000) (000) (000)

Female -004 011 014 005 -004 -001(008) (005) (009) (003) (005) (001)

Born in the UK -024 026 -023 008 -003 -004(016) (015) (015) (009) (008) (003)

Arranged Marriage 011 -008 009 020 -006 001(006) (007) (007) (007) (005) (003)

Discrimination -009 004 001 001 000 000(007) (005) (008) (004) (005) (001)

Children 002 002 -001 001 -001 000(001) (002) (001) (001) (001) (000)

Years Since Arrival -001 001 000 000 000 000(000) (000) (000) (000) (000) (000)

No British Education 001 -014 -006 004 008 001(012) (009) (013) (005) (007) (001)

British Basic Education 005 -012 -013 -009 000 000(014) (010) (017) (006) (007) (001)

British Higher Education -016 -013 -019 001 -002 000(018) (008) (015) (006) (008) (001)

Foreign Education 000 005 -015 -001 001 -002(007) (007) (007) (005) (005) (002)

Employee -014 -004 001 002 -004 -004(010) (008) (010) (005) (006) (002)

Manager -029 -006 009 -001 -017 -003(024) (012) (022) (007) (008) (002)

26

Table 5 Table 4 continued Regression Results for Muslims 330 and non-Muslims 488 to be compared with Table 2 in Bisinet al (2008) Heteroskedasticity corrected (HC1) Standard Errors are in parentheses P-values lt 005 are marked with

Importance of Inter Ethnic Ethnic CompositionReligion Marriage of Schools

Muslims nonndashMuslims Muslims nonndashMuslims Muslims nonndashMuslims

Self Employed 000 -006 011 006 -004 -005(013) (010) (015) (007) (009) (003)

Unemployed -007 007 012 007 -007 -004(008) (010) (010) (008) (007) (002)

No Parents -007 009 004 006 009 000(005) (005) (007) (004) (003) (001)

Parents Physical Contacts 000 000 000 000 000 000(000) (000) (001) (000) (000) (000)

Parents Telephone Calls 000 000 -001 000 000 000(000) (000) (000) (000) (000) (000)

Parents Letters -002 007 003 001 -001 000(003) (003) (003) (002) (002) (000)

English Spoken at Home with Older -034 -008 -005 -009 000 -005(012) (007) (010) (007) (006) (003)

English Spoken at Home with Younger -001 -020 011 003 002 002(007) (009) (008) (009) (005) (003)

English Spoken at Work -004 -019 -008 003 010 004(008) (013) (014) (012) (006) (003)

English Spoken With Friends -004 006 -012 -016 -002 002(007) (009) (010) (009) (007) (002)

Household Income 000 000 000 000 000 000(000) (000) (000) (000) (000) (000)

Discrimination Own Ethnicity 026 013 007 -028 010 020(041) (039) (043) (031) (024) (010)

Ward Density Own Ethnicity 013 -041 -001 036 010 -006(032) (029) (030) (024) (019) (006)

Ward Unemployment Rate 001 001 000 000 000 000(001) (001) (001) (000) (000) (000)

Adjusted R-square 020 009 018 018 005 008

NOTE All estimated models include 7 UK-region dummies and the variables DONOTSPEAKWITHOLDER DONOTSPEAKATWORK andDONOTSPEAKWITHFRIENDS It turned out that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and cannot beincluded in the model

27

131313131313

SULCIS is a multi-disciplinary research center focusing on migration and integration funded

by a Linnaeus Grant from the Swedish Research Council (VR) SULCIS consists of affiliated

researchers at the Department of Criminology the Department of Economics the Department

of Human Geography the Department of Sociology and the Swedish Institute for Social

Research (SOFI) For more information see our website wwwsusesulcis

13

forallexist13ampni()13lowast+13 forall13lowastminus1313131313minus1313

0131ni213132ni3ni(13minus131313minus133 forallexistforall134lowast13

5 6ni7013)ni(88minus13minus133amp13+9minus136exist)lowast13413

6ni113existni(forall1313136exist3

lt 136ampni(forall1313413 =minuslowast1313=13

gt 6ni113existni7013)ni(forall131313minus13forall1313

existni(=cong13139foralllowast13131313131313lowast

Α 13existni0131ni(minus13minus1313131313minus13forall1313131313413

Α Β6ni(minus1313minus13minus13 6exist+4+1313forall134(8

Α forallexistexistni6ni(413exist13minus1381313 )13813

Α 5 13ni(61313413ni413amp13amp4131138

Α 13ni13=13)ni(=Β6813139Α lt =13)ni1313ni(8forall4

13exist9Χ forallexistni∆132ni1313existni(313+)13 forall

lowast13(forallexist13131313139

Χ forallexistni∆132ni1313existni(313+)13 foralllowast13(forallexist13131313139131313ni

Table 2 Weighted and Unweighted Means for Muslims and nonndashMuslims before and after removal of NA compared with Bisinet al (2008)

Muslim Non-Muslim

W e i g h t e d U n w e i g h t e d W e i g h t e d U n w e i g h t e dBefore After Before After Bisin Before After Before After Bisin

et al et al

Importance of Religion 079 081 080 085 079 049 048 048 048 042Attitude Towards Inter Marriage 045 039 045 039 070 019 015 021 017 037Importance of Racial Composition in Schools 006 008 006 007 065 002 001 002 002 033Age at Arrival 2249 2073 2236 2032 3918 2231 1785 2283 1871 4257Age 4042 4201 4018 4056 4440 4398 4490 4459Female 050 047 050 045 047 052 053 052 052 048Born in the UK 008 008 006 006 021 013 018 012 016 028Arranged Marriage 065 063 065 063 022 027 018 030 022 012Discrimination 023 020 020 016 017 038 041 035 039 019Children 312 341 323 343 217 237 232 244 236 168Years Since Arrival 1978 1927 1899 1879 2643 2419 2068 2406 2095 2708No British Education 080 079 083 082 081 056 050 059 053 052British Basic Education 014 013 012 011 006 030 036 029 035 013British Higher Education 005 006 003 003 008 010 012 009 012 016Foreign Education 025 025 023 022 025 026 024 028 025 029Employee 019 021 019 022 038 040 042 040 042 059Manager 002 002 002 002 002 007 008 006 007 004

24

Table 3 Table 2 continued Weighted and Unweighted Means for Muslims and nonndashMuslims before and after removal of NAcompared with Bisin et al (2008)

Muslim Non-Muslim

W e i g h t e d U n w e i g h t e d W e i g h t e d U n w e i g h t e dBefore After Before After Bisin Before After Before After Bisin

et al et al

Self Employed 009 008 007 006 009 012 012 010 010 014OUT OF LABOUR FORCE 050 049 052 048 033 029 036 032Unemployed 019 019 020 022 019 008 010 008 009 008No Parents 065 070 067 070 034 064 065 066 066 032Parents Physical Contacts 269 272 280 288 305 275 302 283 297 387Parents Telephone Calls 210 251 228 285 338 334 334 315 326 474Parents Letters 047 041 048 042 067 025 024 025 027 037English Spoken at Home with Older 012 012 010 009 003 043 053 042 052 008DO NOT SPEAK WITH OLDER 002 004 001 002 002 001 002 002English Spoken at Home with Younger 046 046 041 040 020 077 081 075 079 025DO NOT SPEAK WITH YOUNGER 000 000 000 000 000 000 000 000English Spoken at Work 046 053 041 047 019 079 085 077 082 027DO NOT SPEAK AT WORK 045 039 048 044 016 011 018 014English Spoken With Friends 050 055 046 051 022 079 086 077 083 027DO NOT SPEAK WITH FRIENDS 012 007 012 006 004 002 005 002Household Income 21619 21548 19540 19494 20074 33466 33383 30769 31346 33026Ward Density Own Ethnicity 013 013 016 016 015 011 010 012 011 011Ward Unemployment Rate 1607 1572 1797 1805 1657 1219 1235 1344 1352 1260Discrimination Own Ethnicity 023 022 021 019 021 034 037 034 036 018

25

Table 4 Regression Results for Muslims 330 and non-Muslims 488 to be compared with Table 2 in Bisin et al (2008)Heteroskedasticity corrected (HC1) Standard Errors are in parentheses P-values lt 005 are marked with

Importance of Inter Ethnic Ethnic CompositionReligion Marriage of Schools

Muslims nonndashMuslims Muslims nonndashMuslims Muslims nonndashMuslims

(Intercept) 092 094 083 027 004 -002(030) (026) (037) (026) (018) (004)

Age at Arrival 000 001 000 000 000 000(000) (000) (000) (000) (000) (000)

Female -004 011 014 005 -004 -001(008) (005) (009) (003) (005) (001)

Born in the UK -024 026 -023 008 -003 -004(016) (015) (015) (009) (008) (003)

Arranged Marriage 011 -008 009 020 -006 001(006) (007) (007) (007) (005) (003)

Discrimination -009 004 001 001 000 000(007) (005) (008) (004) (005) (001)

Children 002 002 -001 001 -001 000(001) (002) (001) (001) (001) (000)

Years Since Arrival -001 001 000 000 000 000(000) (000) (000) (000) (000) (000)

No British Education 001 -014 -006 004 008 001(012) (009) (013) (005) (007) (001)

British Basic Education 005 -012 -013 -009 000 000(014) (010) (017) (006) (007) (001)

British Higher Education -016 -013 -019 001 -002 000(018) (008) (015) (006) (008) (001)

Foreign Education 000 005 -015 -001 001 -002(007) (007) (007) (005) (005) (002)

Employee -014 -004 001 002 -004 -004(010) (008) (010) (005) (006) (002)

Manager -029 -006 009 -001 -017 -003(024) (012) (022) (007) (008) (002)

26

Table 5 Table 4 continued Regression Results for Muslims 330 and non-Muslims 488 to be compared with Table 2 in Bisinet al (2008) Heteroskedasticity corrected (HC1) Standard Errors are in parentheses P-values lt 005 are marked with

Importance of Inter Ethnic Ethnic CompositionReligion Marriage of Schools

Muslims nonndashMuslims Muslims nonndashMuslims Muslims nonndashMuslims

Self Employed 000 -006 011 006 -004 -005(013) (010) (015) (007) (009) (003)

Unemployed -007 007 012 007 -007 -004(008) (010) (010) (008) (007) (002)

No Parents -007 009 004 006 009 000(005) (005) (007) (004) (003) (001)

Parents Physical Contacts 000 000 000 000 000 000(000) (000) (001) (000) (000) (000)

Parents Telephone Calls 000 000 -001 000 000 000(000) (000) (000) (000) (000) (000)

Parents Letters -002 007 003 001 -001 000(003) (003) (003) (002) (002) (000)

English Spoken at Home with Older -034 -008 -005 -009 000 -005(012) (007) (010) (007) (006) (003)

English Spoken at Home with Younger -001 -020 011 003 002 002(007) (009) (008) (009) (005) (003)

English Spoken at Work -004 -019 -008 003 010 004(008) (013) (014) (012) (006) (003)

English Spoken With Friends -004 006 -012 -016 -002 002(007) (009) (010) (009) (007) (002)

Household Income 000 000 000 000 000 000(000) (000) (000) (000) (000) (000)

Discrimination Own Ethnicity 026 013 007 -028 010 020(041) (039) (043) (031) (024) (010)

Ward Density Own Ethnicity 013 -041 -001 036 010 -006(032) (029) (030) (024) (019) (006)

Ward Unemployment Rate 001 001 000 000 000 000(001) (001) (001) (000) (000) (000)

Adjusted R-square 020 009 018 018 005 008

NOTE All estimated models include 7 UK-region dummies and the variables DONOTSPEAKWITHOLDER DONOTSPEAKATWORK andDONOTSPEAKWITHFRIENDS It turned out that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and cannot beincluded in the model

27

131313131313

SULCIS is a multi-disciplinary research center focusing on migration and integration funded

by a Linnaeus Grant from the Swedish Research Council (VR) SULCIS consists of affiliated

researchers at the Department of Criminology the Department of Economics the Department

of Human Geography the Department of Sociology and the Swedish Institute for Social

Research (SOFI) For more information see our website wwwsusesulcis

13

forallexist13ampni()13lowast+13 forall13lowastminus1313131313minus1313

0131ni213132ni3ni(13minus131313minus133 forallexistforall134lowast13

5 6ni7013)ni(88minus13minus133amp13+9minus136exist)lowast13413

6ni113existni(forall1313136exist3

lt 136ampni(forall1313413 =minuslowast1313=13

gt 6ni113existni7013)ni(forall131313minus13forall1313

existni(=cong13139foralllowast13131313131313lowast

Α 13existni0131ni(minus13minus1313131313minus13forall1313131313413

Α Β6ni(minus1313minus13minus13 6exist+4+1313forall134(8

Α forallexistexistni6ni(413exist13minus1381313 )13813

Α 5 13ni(61313413ni413amp13amp4131138

Α 13ni13=13)ni(=Β6813139Α lt =13)ni1313ni(8forall4

13exist9Χ forallexistni∆132ni1313existni(313+)13 forall

lowast13(forallexist13131313139

Χ forallexistni∆132ni1313existni(313+)13 foralllowast13(forallexist13131313139131313ni

Table 3 Table 2 continued Weighted and Unweighted Means for Muslims and nonndashMuslims before and after removal of NAcompared with Bisin et al (2008)

Muslim Non-Muslim

W e i g h t e d U n w e i g h t e d W e i g h t e d U n w e i g h t e dBefore After Before After Bisin Before After Before After Bisin

et al et al

Self Employed 009 008 007 006 009 012 012 010 010 014OUT OF LABOUR FORCE 050 049 052 048 033 029 036 032Unemployed 019 019 020 022 019 008 010 008 009 008No Parents 065 070 067 070 034 064 065 066 066 032Parents Physical Contacts 269 272 280 288 305 275 302 283 297 387Parents Telephone Calls 210 251 228 285 338 334 334 315 326 474Parents Letters 047 041 048 042 067 025 024 025 027 037English Spoken at Home with Older 012 012 010 009 003 043 053 042 052 008DO NOT SPEAK WITH OLDER 002 004 001 002 002 001 002 002English Spoken at Home with Younger 046 046 041 040 020 077 081 075 079 025DO NOT SPEAK WITH YOUNGER 000 000 000 000 000 000 000 000English Spoken at Work 046 053 041 047 019 079 085 077 082 027DO NOT SPEAK AT WORK 045 039 048 044 016 011 018 014English Spoken With Friends 050 055 046 051 022 079 086 077 083 027DO NOT SPEAK WITH FRIENDS 012 007 012 006 004 002 005 002Household Income 21619 21548 19540 19494 20074 33466 33383 30769 31346 33026Ward Density Own Ethnicity 013 013 016 016 015 011 010 012 011 011Ward Unemployment Rate 1607 1572 1797 1805 1657 1219 1235 1344 1352 1260Discrimination Own Ethnicity 023 022 021 019 021 034 037 034 036 018

25

Table 4 Regression Results for Muslims 330 and non-Muslims 488 to be compared with Table 2 in Bisin et al (2008)Heteroskedasticity corrected (HC1) Standard Errors are in parentheses P-values lt 005 are marked with

Importance of Inter Ethnic Ethnic CompositionReligion Marriage of Schools

Muslims nonndashMuslims Muslims nonndashMuslims Muslims nonndashMuslims

(Intercept) 092 094 083 027 004 -002(030) (026) (037) (026) (018) (004)

Age at Arrival 000 001 000 000 000 000(000) (000) (000) (000) (000) (000)

Female -004 011 014 005 -004 -001(008) (005) (009) (003) (005) (001)

Born in the UK -024 026 -023 008 -003 -004(016) (015) (015) (009) (008) (003)

Arranged Marriage 011 -008 009 020 -006 001(006) (007) (007) (007) (005) (003)

Discrimination -009 004 001 001 000 000(007) (005) (008) (004) (005) (001)

Children 002 002 -001 001 -001 000(001) (002) (001) (001) (001) (000)

Years Since Arrival -001 001 000 000 000 000(000) (000) (000) (000) (000) (000)

No British Education 001 -014 -006 004 008 001(012) (009) (013) (005) (007) (001)

British Basic Education 005 -012 -013 -009 000 000(014) (010) (017) (006) (007) (001)

British Higher Education -016 -013 -019 001 -002 000(018) (008) (015) (006) (008) (001)

Foreign Education 000 005 -015 -001 001 -002(007) (007) (007) (005) (005) (002)

Employee -014 -004 001 002 -004 -004(010) (008) (010) (005) (006) (002)

Manager -029 -006 009 -001 -017 -003(024) (012) (022) (007) (008) (002)

26

Table 5 Table 4 continued Regression Results for Muslims 330 and non-Muslims 488 to be compared with Table 2 in Bisinet al (2008) Heteroskedasticity corrected (HC1) Standard Errors are in parentheses P-values lt 005 are marked with

Importance of Inter Ethnic Ethnic CompositionReligion Marriage of Schools

Muslims nonndashMuslims Muslims nonndashMuslims Muslims nonndashMuslims

Self Employed 000 -006 011 006 -004 -005(013) (010) (015) (007) (009) (003)

Unemployed -007 007 012 007 -007 -004(008) (010) (010) (008) (007) (002)

No Parents -007 009 004 006 009 000(005) (005) (007) (004) (003) (001)

Parents Physical Contacts 000 000 000 000 000 000(000) (000) (001) (000) (000) (000)

Parents Telephone Calls 000 000 -001 000 000 000(000) (000) (000) (000) (000) (000)

Parents Letters -002 007 003 001 -001 000(003) (003) (003) (002) (002) (000)

English Spoken at Home with Older -034 -008 -005 -009 000 -005(012) (007) (010) (007) (006) (003)

English Spoken at Home with Younger -001 -020 011 003 002 002(007) (009) (008) (009) (005) (003)

English Spoken at Work -004 -019 -008 003 010 004(008) (013) (014) (012) (006) (003)

English Spoken With Friends -004 006 -012 -016 -002 002(007) (009) (010) (009) (007) (002)

Household Income 000 000 000 000 000 000(000) (000) (000) (000) (000) (000)

Discrimination Own Ethnicity 026 013 007 -028 010 020(041) (039) (043) (031) (024) (010)

Ward Density Own Ethnicity 013 -041 -001 036 010 -006(032) (029) (030) (024) (019) (006)

Ward Unemployment Rate 001 001 000 000 000 000(001) (001) (001) (000) (000) (000)

Adjusted R-square 020 009 018 018 005 008

NOTE All estimated models include 7 UK-region dummies and the variables DONOTSPEAKWITHOLDER DONOTSPEAKATWORK andDONOTSPEAKWITHFRIENDS It turned out that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and cannot beincluded in the model

27

131313131313

SULCIS is a multi-disciplinary research center focusing on migration and integration funded

by a Linnaeus Grant from the Swedish Research Council (VR) SULCIS consists of affiliated

researchers at the Department of Criminology the Department of Economics the Department

of Human Geography the Department of Sociology and the Swedish Institute for Social

Research (SOFI) For more information see our website wwwsusesulcis

13

forallexist13ampni()13lowast+13 forall13lowastminus1313131313minus1313

0131ni213132ni3ni(13minus131313minus133 forallexistforall134lowast13

5 6ni7013)ni(88minus13minus133amp13+9minus136exist)lowast13413

6ni113existni(forall1313136exist3

lt 136ampni(forall1313413 =minuslowast1313=13

gt 6ni113existni7013)ni(forall131313minus13forall1313

existni(=cong13139foralllowast13131313131313lowast

Α 13existni0131ni(minus13minus1313131313minus13forall1313131313413

Α Β6ni(minus1313minus13minus13 6exist+4+1313forall134(8

Α forallexistexistni6ni(413exist13minus1381313 )13813

Α 5 13ni(61313413ni413amp13amp4131138

Α 13ni13=13)ni(=Β6813139Α lt =13)ni1313ni(8forall4

13exist9Χ forallexistni∆132ni1313existni(313+)13 forall

lowast13(forallexist13131313139

Χ forallexistni∆132ni1313existni(313+)13 foralllowast13(forallexist13131313139131313ni

Table 4 Regression Results for Muslims 330 and non-Muslims 488 to be compared with Table 2 in Bisin et al (2008)Heteroskedasticity corrected (HC1) Standard Errors are in parentheses P-values lt 005 are marked with

Importance of Inter Ethnic Ethnic CompositionReligion Marriage of Schools

Muslims nonndashMuslims Muslims nonndashMuslims Muslims nonndashMuslims

(Intercept) 092 094 083 027 004 -002(030) (026) (037) (026) (018) (004)

Age at Arrival 000 001 000 000 000 000(000) (000) (000) (000) (000) (000)

Female -004 011 014 005 -004 -001(008) (005) (009) (003) (005) (001)

Born in the UK -024 026 -023 008 -003 -004(016) (015) (015) (009) (008) (003)

Arranged Marriage 011 -008 009 020 -006 001(006) (007) (007) (007) (005) (003)

Discrimination -009 004 001 001 000 000(007) (005) (008) (004) (005) (001)

Children 002 002 -001 001 -001 000(001) (002) (001) (001) (001) (000)

Years Since Arrival -001 001 000 000 000 000(000) (000) (000) (000) (000) (000)

No British Education 001 -014 -006 004 008 001(012) (009) (013) (005) (007) (001)

British Basic Education 005 -012 -013 -009 000 000(014) (010) (017) (006) (007) (001)

British Higher Education -016 -013 -019 001 -002 000(018) (008) (015) (006) (008) (001)

Foreign Education 000 005 -015 -001 001 -002(007) (007) (007) (005) (005) (002)

Employee -014 -004 001 002 -004 -004(010) (008) (010) (005) (006) (002)

Manager -029 -006 009 -001 -017 -003(024) (012) (022) (007) (008) (002)

26

Table 5 Table 4 continued Regression Results for Muslims 330 and non-Muslims 488 to be compared with Table 2 in Bisinet al (2008) Heteroskedasticity corrected (HC1) Standard Errors are in parentheses P-values lt 005 are marked with

Importance of Inter Ethnic Ethnic CompositionReligion Marriage of Schools

Muslims nonndashMuslims Muslims nonndashMuslims Muslims nonndashMuslims

Self Employed 000 -006 011 006 -004 -005(013) (010) (015) (007) (009) (003)

Unemployed -007 007 012 007 -007 -004(008) (010) (010) (008) (007) (002)

No Parents -007 009 004 006 009 000(005) (005) (007) (004) (003) (001)

Parents Physical Contacts 000 000 000 000 000 000(000) (000) (001) (000) (000) (000)

Parents Telephone Calls 000 000 -001 000 000 000(000) (000) (000) (000) (000) (000)

Parents Letters -002 007 003 001 -001 000(003) (003) (003) (002) (002) (000)

English Spoken at Home with Older -034 -008 -005 -009 000 -005(012) (007) (010) (007) (006) (003)

English Spoken at Home with Younger -001 -020 011 003 002 002(007) (009) (008) (009) (005) (003)

English Spoken at Work -004 -019 -008 003 010 004(008) (013) (014) (012) (006) (003)

English Spoken With Friends -004 006 -012 -016 -002 002(007) (009) (010) (009) (007) (002)

Household Income 000 000 000 000 000 000(000) (000) (000) (000) (000) (000)

Discrimination Own Ethnicity 026 013 007 -028 010 020(041) (039) (043) (031) (024) (010)

Ward Density Own Ethnicity 013 -041 -001 036 010 -006(032) (029) (030) (024) (019) (006)

Ward Unemployment Rate 001 001 000 000 000 000(001) (001) (001) (000) (000) (000)

Adjusted R-square 020 009 018 018 005 008

NOTE All estimated models include 7 UK-region dummies and the variables DONOTSPEAKWITHOLDER DONOTSPEAKATWORK andDONOTSPEAKWITHFRIENDS It turned out that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and cannot beincluded in the model

27

131313131313

SULCIS is a multi-disciplinary research center focusing on migration and integration funded

by a Linnaeus Grant from the Swedish Research Council (VR) SULCIS consists of affiliated

researchers at the Department of Criminology the Department of Economics the Department

of Human Geography the Department of Sociology and the Swedish Institute for Social

Research (SOFI) For more information see our website wwwsusesulcis

13

forallexist13ampni()13lowast+13 forall13lowastminus1313131313minus1313

0131ni213132ni3ni(13minus131313minus133 forallexistforall134lowast13

5 6ni7013)ni(88minus13minus133amp13+9minus136exist)lowast13413

6ni113existni(forall1313136exist3

lt 136ampni(forall1313413 =minuslowast1313=13

gt 6ni113existni7013)ni(forall131313minus13forall1313

existni(=cong13139foralllowast13131313131313lowast

Α 13existni0131ni(minus13minus1313131313minus13forall1313131313413

Α Β6ni(minus1313minus13minus13 6exist+4+1313forall134(8

Α forallexistexistni6ni(413exist13minus1381313 )13813

Α 5 13ni(61313413ni413amp13amp4131138

Α 13ni13=13)ni(=Β6813139Α lt =13)ni1313ni(8forall4

13exist9Χ forallexistni∆132ni1313existni(313+)13 forall

lowast13(forallexist13131313139

Χ forallexistni∆132ni1313existni(313+)13 foralllowast13(forallexist13131313139131313ni

Table 5 Table 4 continued Regression Results for Muslims 330 and non-Muslims 488 to be compared with Table 2 in Bisinet al (2008) Heteroskedasticity corrected (HC1) Standard Errors are in parentheses P-values lt 005 are marked with

Importance of Inter Ethnic Ethnic CompositionReligion Marriage of Schools

Muslims nonndashMuslims Muslims nonndashMuslims Muslims nonndashMuslims

Self Employed 000 -006 011 006 -004 -005(013) (010) (015) (007) (009) (003)

Unemployed -007 007 012 007 -007 -004(008) (010) (010) (008) (007) (002)

No Parents -007 009 004 006 009 000(005) (005) (007) (004) (003) (001)

Parents Physical Contacts 000 000 000 000 000 000(000) (000) (001) (000) (000) (000)

Parents Telephone Calls 000 000 -001 000 000 000(000) (000) (000) (000) (000) (000)

Parents Letters -002 007 003 001 -001 000(003) (003) (003) (002) (002) (000)

English Spoken at Home with Older -034 -008 -005 -009 000 -005(012) (007) (010) (007) (006) (003)

English Spoken at Home with Younger -001 -020 011 003 002 002(007) (009) (008) (009) (005) (003)

English Spoken at Work -004 -019 -008 003 010 004(008) (013) (014) (012) (006) (003)

English Spoken With Friends -004 006 -012 -016 -002 002(007) (009) (010) (009) (007) (002)

Household Income 000 000 000 000 000 000(000) (000) (000) (000) (000) (000)

Discrimination Own Ethnicity 026 013 007 -028 010 020(041) (039) (043) (031) (024) (010)

Ward Density Own Ethnicity 013 -041 -001 036 010 -006(032) (029) (030) (024) (019) (006)

Ward Unemployment Rate 001 001 000 000 000 000(001) (001) (001) (000) (000) (000)

Adjusted R-square 020 009 018 018 005 008

NOTE All estimated models include 7 UK-region dummies and the variables DONOTSPEAKWITHOLDER DONOTSPEAKATWORK andDONOTSPEAKWITHFRIENDS It turned out that the variable DONOTSPEAKWITHYOUNGER is TRUE for few observations and cannot beincluded in the model

27

131313131313

SULCIS is a multi-disciplinary research center focusing on migration and integration funded

by a Linnaeus Grant from the Swedish Research Council (VR) SULCIS consists of affiliated

researchers at the Department of Criminology the Department of Economics the Department

of Human Geography the Department of Sociology and the Swedish Institute for Social

Research (SOFI) For more information see our website wwwsusesulcis

13

forallexist13ampni()13lowast+13 forall13lowastminus1313131313minus1313

0131ni213132ni3ni(13minus131313minus133 forallexistforall134lowast13

5 6ni7013)ni(88minus13minus133amp13+9minus136exist)lowast13413

6ni113existni(forall1313136exist3

lt 136ampni(forall1313413 =minuslowast1313=13

gt 6ni113existni7013)ni(forall131313minus13forall1313

existni(=cong13139foralllowast13131313131313lowast

Α 13existni0131ni(minus13minus1313131313minus13forall1313131313413

Α Β6ni(minus1313minus13minus13 6exist+4+1313forall134(8

Α forallexistexistni6ni(413exist13minus1381313 )13813

Α 5 13ni(61313413ni413amp13amp4131138

Α 13ni13=13)ni(=Β6813139Α lt =13)ni1313ni(8forall4

13exist9Χ forallexistni∆132ni1313existni(313+)13 forall

lowast13(forallexist13131313139

Χ forallexistni∆132ni1313existni(313+)13 foralllowast13(forallexist13131313139131313ni

131313131313

SULCIS is a multi-disciplinary research center focusing on migration and integration funded

by a Linnaeus Grant from the Swedish Research Council (VR) SULCIS consists of affiliated

researchers at the Department of Criminology the Department of Economics the Department

of Human Geography the Department of Sociology and the Swedish Institute for Social

Research (SOFI) For more information see our website wwwsusesulcis

13

forallexist13ampni()13lowast+13 forall13lowastminus1313131313minus1313

0131ni213132ni3ni(13minus131313minus133 forallexistforall134lowast13

5 6ni7013)ni(88minus13minus133amp13+9minus136exist)lowast13413

6ni113existni(forall1313136exist3

lt 136ampni(forall1313413 =minuslowast1313=13

gt 6ni113existni7013)ni(forall131313minus13forall1313

existni(=cong13139foralllowast13131313131313lowast

Α 13existni0131ni(minus13minus1313131313minus13forall1313131313413

Α Β6ni(minus1313minus13minus13 6exist+4+1313forall134(8

Α forallexistexistni6ni(413exist13minus1381313 )13813

Α 5 13ni(61313413ni413amp13amp4131138

Α 13ni13=13)ni(=Β6813139Α lt =13)ni1313ni(8forall4

13exist9Χ forallexistni∆132ni1313existni(313+)13 forall

lowast13(forallexist13131313139

Χ forallexistni∆132ni1313existni(313+)13 foralllowast13(forallexist13131313139131313ni