Residential mobility in divided societies: How individual religion and geographical context influenced housing moves in Northern Ireland 2001-2011

Residential segregation is pervasive in many societies. People making residential moves in these divided contexts may increase or decrease segregation levels. In this paper, the divided society of Northern Ireland is used as an example to explore how residential mobility relates to residential segregation by religion. Survey evidence for this country consistently shows a preference for mixed neighbourhoods, yet actual patterns of geographical mobility suggest people move to same ‐ religion areas. The paper uses the Northern Ireland Longitudinal Study (NILS) to explore the individual and contextual factors that influence the destinations of internal migrants by religion between 2001 and 2011. How they move up or down the contextual ‘ religion ladder ’ of localities is modelled with reference to both their individual socio ‐ demographic and neighbourhood characteristics in 2001. It is found that there are still individual religious differentials in people's destinations. Catholics, for instance, are more likely than Protestants to move to more Catholic areas, suggesting that individual religion remains important despite the Peace Process. Some possible reasons for this are considered with a partial explanation being found in the geographical patterning of the population. Existing patterns of residential segregation constrain moves in religious space for the majority of people. It is concluded, nevertheless, that an individual's religion remains a considerable factor contrary to expectations.

dimensions, such as isolation, exposure and unevenness (Massey & Denton, 1988;Peach, 2009). Furthermore, it can be conceptualised and measured as a global statistic for an entire area or as a local statistic for different places (see Lloyd, Shuttleworth, & Wong, 2014). This paper uses the example of Northern Ireland to investigate how individual sociodemographic and neighbourhood characteristics shape residential moves through religiously segregated space; and the latter is conceptualised as the religious composition of an area measured in terms of the proportionate preponderance of either Catholics or Protestants.
Northern Ireland is an interesting case because in most respects, it is similar to many other western societies; the housing market today operates without overt discrimination although social class and income are constraints on accessing housing opportunities. It differs, however, in being a divided society with a history of political territorialisation. The major divide is religion. This is invisible at an individual level (unless someone wants to make their identity clear) unlike the often-visible marker of race, as in the United States or the remainder of the United Kingdom, although territories, schools, churches and places are very clearly associated with religious communities and identities (e.g., flags, curb markings and murals). Northern Ireland remains residentially segregated by religion and is therefore a country in which residential mobility occurs in an already highly structured geographical context (Shuttleworth & Lloyd, 2009). On the other hand, the level of overt sectarian violence diminished noticeably during the 1990s, a process that was reinforced by the signing of the Belfast/Good Friday agreement in 1998. Moreover, as we will show below, subsequent surveys in Northern Ireland indicate that the vast majority of the Northern Ireland population support the idea of mixed communities, whereas nonsectarian factors like house size and affordability seem to be just as important in migration decision making there as in other countries.
The main aim of this paper, therefore, is to examine whether, in the post-Troubles environment of Northern Ireland, religion is still playing a major role in influencing people's residential-mobility behaviour. Specifically, it addresses three related questions: (1) Is there still any difference between Catholics and Protestants in the geography of their residential mobility? (2) Is residential mobility acting to increase or decrease segregation by religion? (3) How important is geographical context in shaping the residential mobility of these two religious groups? These questions are addressed using the Northern Ireland Longitudinal Study (NILS) which provides individual-level data on change of address between the 2001 and 2011 Population Censuses (Dibben, Shuttleworth, Shelton, & Duke-Williams, 2018), allowing us to measure the net effects of 10 years of address changing on the distribution of these two groups and relate people's home moves to the structuration of religious space at neighbourhood level.
The paper starts by considering the general relationship between residential segregation and residential mobility and how religion might influence residential choice. It then outlines the Northern Ireland context, discussing the Peace Process, the policy background, the path followed from conflict and the likely implications for residential segregation. After defining specific research questions, the paper then describes our NILS dataset and the methods used for the primary analysis. The results are then presented and discussed, with a focus on the individual and geographical factors that shape moves through religiously segregated residential space and the consequences for better understanding society and assessing the relevance of policies designed to reduce segregation.

| RESIDENTIAL SEGREGATION AND RESIDENTIAL MOBILITY: SOME WIDER CONSIDERATIONS
The wider literature is useful in considering how religion influences where people live and their destinations when they change address.
This allocation is nonrandom. People tend to move to neighbourhoods which share their personal characteristics. Migration, whether in terms of social deprivation or ethnicity, puts people in their place (Bailey, 2012). There are a number of contrasting ways in which the allocation of individuals and households to different places can be conceptualised and explained, with different levels of individual agency.
The first of these interpretations is self-segregation, which tends to stress individual agency, where groups voluntarily live apart. This has been a dominant political explanation, especially of the segregation of South Asian populations in Britain, although it has been subject to robust challenge (Coulter & Clark, 2019). This raises the question of why self-segregation takes place and why ethnic enclaves (Lersch, 2012) exist. One reason might be externalities that arise from segregation, such as strong mutual support networks and the availability of community resources and facilities (such as religious buildings and schools). Another, less benign, interpretation is in terms of defensive space and fear. This has been one interpretation of residential segregation in Northern Ireland in the context of the political significance of territory (Anderson & Shuttleworth, 1998;Boal, 2002). This can be elaborated by the spatially limited perceptual maps of residents (Green, Shuttleworth, & Lavery, 2005) which mean that individuals are unaware of the full range of employment and housing opportunities. This might be most severe in politically divided societies like Northern Ireland but is typical of even 'normal societies'.
Other interpretations tend to place less emphasis on individual agency and more on structural constraints. Although individuals may be theoretically free to choose where they live, there are limitations imposed by their access to resources. Thus, in the spatial assimilation model, individuals have preferences to locate in places alongside the same type of people with two or more population groups having the same preferences (Lersch, 2012). This might be in terms of lifecycle, ethnicity or income. However, some people or groups may be unable to realise their preferences given wealth and income constraints (South & Crowder, 1997), especially given the structural inequalities that lead to group differences. Moving further on the road to structural constraints is that of place stratification (Lersch, 2012). In this approach, dominant groups try to impose structural constraints on the access of others to neighbourhoods. This might be done through control of social housing (Clark et al., 2014), with bureaucratic gatekeepers, or through less obvious discrimination in the private sector. Which of these approaches, if any, apply to contemporary Northern Ireland? Current patterns of residential segregation are not the result of large-scale ongoing discrimination (although this does not rule out past practice in shaping the present), and there is consociational power sharing between British Unionists and Irish Nationalists.
There is a declining social housing sector and, as is the case of the rest of the United Kingdom (and Ireland), the majority of housing is privately owned, either owner occupied (by far the largest sector) or privately rented.1 This leaves little scope for the exercise of exclusionary power by the state and equality legislation now rules out discriminatory practices. Considering the spatial allocation concept, Northern Ireland has a free-market economy with formal equal rights for all citizens. As elsewhere, the main constraints on access to housing in the private sector are income and availability. Because legislation and political change have led to greater economic equality between Catholics and Protestants, discrimination in the allocation of housing and work, as was the case in the past, (Barritt & Carter, 1962;Smith & Chambers, 1991) has been prohibited. The Catholic proportion of the population has also grown steadily since 1971, and it is unlikely that there are economic and demographic inequalities between these two religious groups that constrain housing choice differentially. This leaves self-segregation and residential preferences by religion. In this case, the evidence is not conclusive. Northern Ireland thus differs from other segregated societies in a number of important ways, limiting the current applicability of some of the concepts used elsewhere (although they may have had more explanatory power in Northern Ireland's more unequal past). It differs in that its main social divide is not characterised by visible physical difference as is the case in the racialised housing market of the United States (Crowder & South, 2008). Settlement patterns are not currently shaped by deliberate acts of policy as in Israel/Palestine, although communal violence in the past, and state policing and planning responses to this, have doubtlessly shaped residential patterns historically (Sheehan & Tomlinson, 2018). It cannot be described either as a society of recent immigration into which newcomers are assimilated as in societies which saw immigration in the late 20th and early 21st Centuries: The religious settlement geography of Northern Ireland dates back to the 17th Century (Gregory, Cunningham, Lloyd, Ell, & Shuttleworth, 2013). Finally, it is not now a society which has major demographic and power asymmetries, where one group is a visible minority, has far fewer economic prospects, and which faces continued discrimination and political exclusion, although it might so have been described in the past. More information on the peculiarities of Northern Ireland are therefore provided in the next section.

| THE NORTHERN IRELAND CONTEXT
The prime sociodemographic division in Northern Ireland as a marker of national/political preference is religion rather than race, language or social class. In this, Northern Ireland falls into the same class as some other divided countries, for example, the successor states of the former Yugoslavia. The two communities that dominate political discourse are Irish Nationalists/Republicans (most often Roman Catholics) and British Unionists/Loyalists (who are mostly Protestant) but identities can be more complex than this and may differ between cultural and political domains with, for example, a Northern Irish identity (Tonge & Gomez, 2015). Nevertheless, the durability of the Protestant/Catholic dichotomy points both to its historical relevance and its current importance in political and public discourse even after the Peace Process. Because the division between people in Northern Ireland is invisible, it needs considerable knowledge to 'tell' (or make an informed guess) about someone's identity (Brewer, 1992). It also requires considerable effort to maintain group boundaries because they are otherwise so small-Freud's narcissism of minor differences.
One way in which boundaries are produced and reproduced is through politicised and sectarianized territories which are associated with each community (Anderson & Shuttleworth, 1998); knowing where someone lives is one of the main diagnostics in 'telling' . The emphasis on territory and residential segregation is not surprising because Northern Ireland was delineated in 1921 in a partition of the island of Ireland that was designed to give it a 'safe' Protestant majority. Population balance and its distribution have been a pressing concern at the level of the state since then but has also cascaded down to smaller territories and places.
The territory of Northern Ireland experienced civil unrest and violence before, during and after partition, with residential segregation increasing through time (Smith & Chambers, 1991). Violence has mainly been episodic with the most recent major outbreak being 'The  (Gregory et al., 2013;Power & Shuttleworth, 1997). These developments also acted to increase residential segregation. This reached its peak and plateaued in the 1991 and 2001 Censuses (Shuttleworth & Lloyd, 2009), with the 2011 Census showing a reversal in this trend .
What is clear is that government policy has seen desegregation as desirable. The 1998 Belfast/Good Friday Agreement legally enshrined equality between the two major Northern Ireland communities and was underpinned by power sharing and the Section 75 provision of the Belfast Agreement to place equal opportunity and good relations centrally in the design, delivery and assessment of public policy (ECNI, 2010). Besides achieving greater equality, the intention was to break down segregation in various domains of life such as the labour market and housing (OFMDFM, 2005). The promotion of mixed and shared communities therefore is important as a policy issue.
The most recent round of policy is "Together Building a United Community" (TBUC). This was launched in 2013 as a way to deliver on previous commitments to good relations (TEO, 2013). It has a multifaceted approach aiming to work across a wide remit which includes education, culture, the labour market and housing. In housing, it aims to improve community relations, reduce segregation and encourage mixing, and it recognises that government intervention in partnership with housing associations and the Housing Executive is required. It aims to preserve neighbourhoods that are already mixed, but in addition, it seeks to build 10 new shared estates, half of which had already

| DATA AND METHODS
For the purposes of this study, it is necessary to have longitudinal data that captures information on address changes between two points in time, as well as data on individual social and demographic characteristics and on the geographical context of places that movers leave and to which they move. All these data are available in the NILS and, indeed, also in the other UK Census Longitudinal Studies-the Scottish Longitudinal Study (SLS) and the Office for National Statistics Longitudinal Study (ONS LS), see Dibben et al. (2018). Accordingly, data were extracted from the NILS in an 'approved project' for this research.
The NILS is built on a spine of health card registration data to which censuses are linked and is based on a sample of around 28% of the Northern Ireland population (drawn from 104 out 365 birthdates) and links data from the 2001 and 2011 Censuses. NILS is well suited to detailed internal migration study because its high sampling fraction allows the use of Super Output Areas (SOAs) as a unit for geographical analysis, areas with around 2,000 individuals and 500 households (Reid, Crone, & Hayes, 2018). Additionally, it is possible to identify all address changes and not just those that cross SOA boundaries.
Another strength is the high census-to-census linkage rate (it has very low attrition unlike survey-based longitudinal studies), but a weakness is the lack of intercensal attribute information. The analytical dataset did not include the entire NILS sample but was restricted to those (see Clark et al., 2014) and over what distances.
Neighbourhood type and geographical context were specified in two contrasting ways. The first, and simplest, was to take the percentage Catholic of the SOA in which a person was living at the start of the reference period in 2001. Secondly, Local Indicators of Spatial Association (LISA) clusters were calculated, using GeoDa™ (Anselin, 1995(Anselin, , 2003 with the percentage Catholic per SOA as input. The clusters were created using the default 12 km distance threshold. This placed each SOA into five spatial classes-random (with no significant local spatial autocorrelation at the 5% level), low, low (a low Catholic SOA surrounded by other low Catholic SOAs), high, low (a highly Catholic SOA surrounded by low Catholic SOAs), low, high (a low percentage Catholic SOA surrounded by highly Catholic SOAs), and high, high (high percentage Catholic SOAs surrounded by highly Catholic SOAs).
Given the inherently hierarchical structure of the data, with NILS members nested in SOAs in 2001, the most appropriate method was multilevel modelling (Goldstein, 2010). This allowed individual characteristics to be included as fixed effects in our models but random variation across all SOAs to be measured and, most importantly, the crosslevel interaction between Level 1 individual variables (such as religion) and Level 2 (SOA variables) such as SOA religious composition (together with a second-order polynomials for the latter) to be evaluated. A hierarchical two-level individuals-in-neighbourhoods model which includes both individual (Level 1) and area specific (Level 2) predictors and their cross-level interaction (Gould, 2010), with random 'intercepts' (i.e., means) allowed to vary between SOAs can be written thus: y ij = β 0 x 0ij + β 1 x 1ij + β n x nij + α 1 w 1j x 0ij + α 2 w 2 1j x 0ij + α 3 w 1j x 1ij + α 4 w 2 1j x 1ij x 0 the constant; x 1 a predictor variable measured at Level 1 (e.g., religion); w 1 A predictor variable measured at Level 2 (e.g., religion of the SOA of origin in 2011); β0 the estimated intercept term; β 1 -β n the estimated slope terms associated with Level 1 predictor variables; α 1 -α n A number of estimated slope terms associated with the Level 2 predictor variable and its cross-level interaction with the Level 1 predictor; ε 0 the Level 1 individual random terms; μ 0 the Level 2 SOA random terms.
All the Level 1 predictors are binary categorical variables (e.g., a dummy variable is used to distinguish Catholics from the reference category Protestant), in the analysis presented below, and that contrast coding is used for model specification (Jones, 1991). The two sets of random terms in Equation 1 can be summarised by two estimated variances σ 2 ε 0 and σ 2 μ 0 and are associated with the constant that has been allowed to vary at both Level 1 and Level 2. Additional to the three research questions identified earlier in the paper, this multilevel analytical framework also permits two related overarching questions to be investigated. Has religion ceased to be important as a factor in determining people changing address changers move from and to, and are there any differences between Protestants and Catholics with respect to how likely they are to move to more or less Catholic contexts?

| RESULTS
In Table 1,

Protestant × not significant (Reference)
-    In considering whether Catholics and Protestants differ in their migration behaviour, the evidence of Table 2 shows that Catholics-and indeed those with other or no religions-were more likely to move to more Catholic places than Protestants. It is important to note that this is so even after controlling for age, housing tenure, highest qualification, economic status and health. In the past differentials between Catholics and Protestants along these dimensions were high (see Osborne & Shuttleworth, 2004;Sheehan & Tomlinson, 2018;Smith & Chambers, 1991) and, even though since the 1990s they have very much lessened (Rowland, 2019), some still remain, with social deprivation on average being higher for Catholics. It is therefore essential to make controls for a wide range of social and demographic statistics so as to isolate, as far as possible, the independent effects of religion on mobility.
For the demographic group studied here in the multivariate How could this be the case? We suggest three potential explanations. One is that people are 'shy sectarians', in that being questioned about a potentially sensitive topic like this they do not admit to preferring neighbourhoods of their own religion. Another is that people may be subject to 'unconscious bias' that influences their locational choices when it comes to considering where to live. A third possibility is that people may want to have mixed neighbourhoods but, without making this the prime factor in the decision making, inevitably end up in roughly the same type of place as the one in which they started because of population structures and the 'normal' spatiality of internal migration. Given that the population geography of Northern Ireland is strongly structured by religion (Lloyd, 2010) and also that most internal migration in Northern Ireland-as elsewhere-are short-distance in nature , it is likely that many movers start and end in the same type of place by religion. As such, they could be considered 'prisoners of spatial structure' . This adds to the evidence base that many address changes start and end in the same general type of place (Clark et al., 2014;Clark & Morrison, 2012), with few individuals spanning the whole area hierarchy.
No doubt a combination of all these conditions is responsible for the wide discrepancy between the NILT survey results and people's actual behaviour. Certainly, our study provides no evidence that supports absolutely one of the three alternative explanations as opposed to the others. What we have found is that an individual's religion remains important in determining the type of area that members of our sample had moved to during the decade leading up to the 2011 Census. Equally, our analyses rule out geographical context as being the only full explanation of the migration patterns observed; there are between-place differences but individual religion remains important.
Overall, therefore, it is the combination of individual and neighbourhood characteristics that shapes movement through religious space.

| IMPLICATIONS
The analysis reported in the paper has wider implications. For social and housing policy in Northern Ireland, it suggests that aspirations for greater residential social mixing are unlikely to be met in the near future by the status quo. As in other parts of the UK, owner occupation is the leading housing sector, followed by private renting (an increasing share since 2001) and then social housing. Most, but not all, housing moves therefore take place in the owner-occupied and private-rented sectors. These operate as a free market, and it thus seems that market-directed moves in these sectors alone are insufficient to change population distributions enough to diminish religious segregation. As regards the socially rented sector, here, there is scope for state intervention to provide new housing stock for mixed-religion tenancy. This can be successful in some circumstances in developing mixed estates (Stevenson et al., 2019) but can also face significant challenges related to geographical context-inner-city locations, often with histories of sectarianism, are unfavourable locations for mixedhousing interventions, for instance-and to paramilitary threats (Belfast Telegraph, 2019). In this regard, it is intriguing that individual religion still seems to be a strong factor, at least up to 2011. This is some 17 years after the first paramilitary ceasefires of 1994, and 13 years after the Belfast/Good Friday Agreement, so considerable time has elapsed but arguably perhaps not enough. The 2021 Census results, when available, will reveal more.
In terms of possible applicability elsewhere besides Northern Ireland, this paper has elaborated a method (Clark & Morrison, 2012) that can be used to examine the interaction between people, places, and internal migration in divided societies, irrespective of whether the main dimension of division is religion, ethnicity or social class. The use of LISA clusters takes the analysis further by considering the role of local population structure in shaping the opportunities and barriers for moves between socially different, or similar, places. Model parameters such as those presented above might also be useful as empirical parameters for use in agent-based models in the tradition of Schelling (1971). Their strength in this case is based not on what people answer in surveys but on what they actually do and, as we have seen, these can be quite different things.