Levels or changes?: Ethnic context, immigration and the UK Independence Party vote

Will the rising share of ethnic minorities in western societies spark a backlash or lead to greater acceptance of diversity? This paper examines this question through the prism of the UK Independence Party (UKIP), the most successful populist right party in British history. The paper contributes to work on contextual effects by arguing that ethnic levels and changes cross-pressure white opinion and voting. It argues that high levels of established ethnic minorities reduce opposition to immigration and support for UKIP among White Britons. Conversely, more rapid ethnic changes increase opposition to immigration and support for UKIP. Longitudinal data demonstrates that these effects are not produced by self-selection. The data further illustrate that with time, diversity levels increase their threat-reducing power while the threatening effects of ethnic change fade. Results suggest that the contextual effects literature needs to routinely unpack levels from changes. This also suggests that if the pace of immigration slows, immigration attitudes should soften and populist right voting decline.


Background
The contact hypothesis, originally developed by social psychologists, argues that when members of the dominant ethnic groupi.e. native-born whites in the US or White British in the UK -have the chance to positively interact with minorities in their locale, they become more comfortable with difference (Allport, 1954;Pettigrew and Tropp, 2006). Diverse contexts tend to lead to more contact. For instance, a White British person has a 0.65 probability of having no minority friends if she lives in a homogeneous ward, but this drops -controlling for individual and area attributes -to only 0.07 in a ward comprised of 90 percent minorities (UKHLS, 2014). The literature thus finds that contextual diversity ameliorates white attitudes to immigration and outgroups, especially in small geographies below that of the ward/census tract (10,000 population). Thus Kaufmann and Harris (2015) locate 24 studies using contextual variables from low geographies and find that three-quarters report a positive contact effect.
Against the contact hypothesis, the threat hypothesis claims that diversity stimulates white opposition to immigration. Work on the racial threat hypothesis, beginning with the landmark studies of Key (1949) and Blalock (1957) revealed higher levels of white segregationist voting in Southern counties with higher proportions of African-Americans, as in the Mississippi Delta or coastal South Carolina. Later work generalized these findings to the Midwest and Northeast (Stein et al., 2000: 286). As Robert Putnam notes, diversity, far from incubating toleration, often results in heightened inter-ethnic conflict (Putnam, 2007).
Beyond the contextual literature, this work is concerned more broadly with the question of political demography e whether population change arising from differences in rates of birth, death, migration or age structures between populations affects politics (Goldstone et al., 2012). It therefore contributes to work on differential ethnic population growth, a subfield of political demography which examines the effect of ethnic change on collective perceptions and behaviour (i.e. Toft, 2007;Fearon and Laitin, 2011).

Levels or changes?
Echoing the political demography perspective, scholars have recently questioned whether variation in levels of minority presence offers the best test of the contact and threat hypotheses (Newman and Velez, 2014). Instead, they counter that the accent should be placed on ethnic change, which induces a sense of dissonance among native-born whites. In a separate project with Matthew Goodwin, I performed an exhaustive meta-analysis of all papers on immigration attitudes or far right voting we could find published between 1995 and early 2016 that included an ethnocontextual independent variable (Kaufmann and Goodwin, 2016). Across the contextual effects literature, we identified 69 tests of ethnic change on immigration opinion or far right voting in 32 published studies since 1995. These appear in Appendix I. 1 52 of 69 tests find that ethnic change at some level of geography is associated with increased opposition to immigration or higher populist right support. When restricted to studies reporting statistically significant contextual effects, 27 of 31 tests (87%) across 17 studies report a threat effect and just 4 a significant contact effect (see Appendix II). This compares to only 65% of 37 statistically significant tests in 26 studies based on ethnic levels which identify a threat effect (35% find a threat reduction effect) (see Appendix V). Virtually none of these control for ethnic change. Thus while contextual studies using minority change as the contextual measure overwhelmingly report threat effects, those based on minority levels are more evenly split between threat and contact outcomes. We also found, using a funnel plot, no evidence of publication bias towards reporting positive results.
To cite just a few examples: Hopkins (2010) finds that nativeborn white Americans in ethnically-changing neighbourhoods, when sensitised by the salience of immigration in the news, are significantly more anti-immigration than those in more demographically stable areas. 2 Harris (2012: 177, 220) uncovers strong positive effects of minority change on the extreme right British National Party (BNP)'s ward-level vote share using ecological analysis. Newman (2013) shows that rapid minority increase in previously non-diverse counties is associated with greater white American hostility to immigration. Abrajano and Hajnal (2015: 132) report that a high rate of Latino growth in a county is associated with a positive stimulus to anti-immigration views. Individuals with authoritarian personality profiles are particularly sensitive to change: their preference for order interacts with local ethnic shifts to produce highly negative attitudes to immigration (Johnston et al., 2015). Qualitative work, meanwhile, finds abundant evidence of white anti-immigration mobilisation in ethnically transforming locales such as Carpentersville, Illinois, Farmer's Branch, Texas or Barking and Dagenham, England (Vicino, 2013;Gest, 2016).
Only two published articles on populist right voting I am aware of report significant effects for both ethnic levels and changes. Kessler and Freeman (2005: 276e9) show that when a term for ethnic change is introduced alongside ethnic levels, the sign for levels switches from positive to negative while ethnic change strengthens its power to predict European populist right party support. This did not form the basis of an argument about the divergent effects of levels and changes. This is analogous to this paper's claim. The article did not, however, elaborate upon this finding. The other study is an article on Front National support in France using ecological regression. It found the opposite relationship to that postulated here at the D epartement (population median of 500,000) level, though not at the smaller Commune level (i.e. Della Posta, 2013).

Hypotheses
From the literature, I hypothesise that local minority levels reduce native-born white threat perceptions while changes elevate it. Given the organic connection between ethnic change and ethnic levels, I also postulate a habituation mechanism: as changes translate into high ethnic levels, however, the effects of the past are transformed. Thus time reduces threat as changes fade and older levels of ethnic diversity gain legitimacy. Hence I expect that: H1. Higher levels of ethnic minorities in a respondent's context predict lower UKIP support and opposition to immigration H2. Faster increase in ethnic minorities in a respondent's context predicts higher UKIP support and opposition to immigration H3. Fading of change over time: faster increase in ethnic minorities in a respondent's context in the 2000s predicts higher UKIP support, but this is less true for ethnic changes from the previous decade (1990s) H4. Enhanced minority establishment over time: higher levels of ethnic minorities in a respondent's context over a decade ago (1990s) are a stronger predictor of lower UKIP support and opposition to immigration than levels of ethnic minorities in the current decade (2000s) Finally, H5. The association between higher minority levels and lower threat is not produced by the self-selection of whites who dislike diversity out of diverse areas leaving disproportionately tolerant whites behind

Data
I use several sources for this research. In order to measure UKIP support and the political profile of White British movers, I use waves 1e4 of the Understanding Society (UKHLS) survey. Understanding Society is an annual longitudinal study of approximately 45,000 individuals, including a minority boost sample. I use 1991, 2001 and 2011 ward-level British census data which is attached to individual survey records in the UKHLS. A common 2001 ward 1 Tests include either: a) separate studies using the variable; b) separate models within studies using the variable; or c) separate terms within models using the variable.
2 Some claim that perceptions of immigrant presence do not match reality, and thus the 'treatment' of immigration is inchoate. Yet, while respondents generally overstate the share of minorities and immigrants in their locale, those in more diverse or immigrant-rich zip codes and counties report much higher levels and changes than those in less diverse places (Newman and Velez, 2014: 6e7).
geography is used to link census data across the three census dates. 3 Note that UK wards average approximately 6500 population so offer a good example of a low-level geography in which I might expect to find contact effects. While similar to the Panel Study of Income Dynamics (PSID) in the United States, UKHLS contains modules covering a wider array of subjective measures. Party vote, political participation, political attitudes, reasons for moving, attitudes to locale and national identity are included in at least some survey waves. This permits the analyst to guard against endogeneity, i.e. whether whites who support UKIP are more likely to leave diverse areas. The UKHLS enables the user to examine the beliefs of whites who leave, enter and remain in diverse areas, enabling the generation of a four-year longitudinal panel dataset of White British incomers to, outmigrants from, and stayers in, diverse or homogeneous wards. The sample consists of 170,460 person-years, of which 124,524 person-years of responses are obtained from approximately 46,500 White British individuals. 83 percent of the wave 2 sample responded in wave 3, though attrition rates are higher among ethnic minorities, youth and movers. 4 In addition, I examine the effect of Local Authority ethnic levels and changes on reported UKIP vote in the 2014 European election using the British Election Study (BES) 2015 Combined Internet Panel Study. Data is drawn from waves 1 and 2 of the BES (Fieldhouse et al., 2015), a sample of over 24,000 individuals across the UK including almost 6000 UKIP voters.
To assess how ward ethnic context affects white attitudes to immigration, I draw on the Home Office Citizenship Surveys, which sample approximately 15,000 respondents per year -5000 from non-European minority groups and 10,000 whites (Office for National Statistics and Home Office, 2010, 2011). The survey was conducted biennally or annually in England and Wales between 2001 and 2011. Ward-level geocoded data for the 2009e10 and 2010-11 surveys have been obtained through survey firms. Other years are not available at the geographic scale required. 5 The pooled 2-year sample yields approximately 16,000 White British respondents, the target group. The survey is rich in questions pertaining to attitudes toward immigration, ethnic relations and perceptions of locality.
A final source is 2010, 2011 and 2012 Local Government election results from the University of Plymouth Elections Centre, which are attached to 2011 ward-level census data. This permits the analysis of a much larger sample of UKIP (and BNP) voters than is possible in surveys such as UKHLS. 6

Method
The analysis proceeds as follows. I begin by examining individual-level UKIP voting from four waves of the UKHLS. This analysis includes a test of endogeneity, i.e. the 'white flight' hypothesis. This is followed by several robustness checks. The first focuses on immigration opinion in the Citizenship Surveys. This followed by an analysis of UKIP voting in the British Election Study (BES) using Local Authority (LA)-level contextual parameters (ward-level parameters are not available for BES due to disclosure risk). Finally I include an ecological analysis of UKIP and British National Party (BNP) voting at ward level for the years 2010e12.

UKIP support in the UKHLS
As noted, I first perform an analysis of support for the UK Independence Party (UKIP) at individual level, using Understanding Society (UKHLS) data.

Dependent variable
The dependent variable for this study is comprised of the combined response to the questions 'If there were to be a general election tomorrow, which political party do you think you would be most likely to support?' and 'which party do you feel closest to?' 7 While party support and vote intention are discrete variables, there is a close relationship between the two. Moreover, the two questions are asked of different respondents in each wave, so amalgamating them does not duplicate responses. One of the response categories in both questions is 'other party', followed by a write-in option. This is where a UKIP or other populist right response may appear. Only 10e14 percent of those who said 'other party' failed to name a recognisable party and most of these were 'don't know' or uncodeable responses and thus are unlikely to be skewing the data. This generates 2091 UKIP person-years out of approximately 130,000 White British person-years of data on the two political questions across waves 1e4.
UKIP support amounts to just 2 percent of the sample (3.1 percent of White British who supplied a political response) in wave 4 [2012e14]. This is a considerable undercount given UKIP's actual 2014 European election popular vote of 26.6 percent and 2015 general election result of 12.6 percent. Wave 4 is largely drawn from 2012 to 13 responses, and while UKIP support was rising in this period from a lower base (16 percent in 2009 European elections, 3 percent in 2010 general election), the structure of the UKHLS partisanship questions underestimates UKIP support and I must be attentive to the possibility this selects for more committed UKIP partisans. Particular individuals do account for multiple person-years of data, yet the data show a notable increase in UKIP support in wave 4 and considerable individual variability. For instance, only around 20 percent of UKIP supporters in a given wave were supporters in the previous wave; for vote intenders, this rises to 35 percent. Among vote intenders in a given wave less than 15 percent were party supporters in a previous wave, and vice-versa. Once again, the dependent variable is a dummy for UKIP support or vote intention in a given year.

Independent variables
Individual-level variables include age, sex, highest educational qualification (run as a continuous variable) and income, which previous analyses suggest would predict a UKIP vote (Ford and Goodwin, 2014). Contextual parameters are drawn from the census (ONS, 2013). These include the proportion of the ward of residence comprised of non-European ethnic minorities and the rate of non-European minority increase in the ward since 2001. Similar shares are calculated for levels and changes in 'White Other' (mainly Eastern European) share and for combined non-European and White Other 'minority' total.

Results
Results of a logistic regression of UKIP vote, with robust standard errors, is presented in Table 1. The first point to notice is that individuals in the most recent wave are associated with a significantly greater likelihood of supporting UKIP. This reflects the rising trajectory of UKIP over the 2012e14 period (Ford and Goodwin, 2014). 8 As expected, older, male, less educated and poorer voters are significantly associated with support for UKIP, reflecting the 'left behind' social profile advanced by Ford and Goodwin.
Focus on the effects of the contextual parameters in the lower half of the table. Notice that the level of minorities in a respondent's ward of residence predicts a significantly lower likelihood of supporting or intending to vote for UKIP (H1). Yet changes in minority share are associated with the obverse: a significantly elevated likelihood of backing UKIP (H2). Levels and changes have disparate effects in the hypothesised (H1 and H2) direction. The change effect for UKIP voting is robust to the substitution of 1991 or 2001 minority share for 2011 minority share. Deprivation and population density in ward were not significant. Local Authority levels and changes are not significant when ward levels or changes are present in the model.
Non-European minority demographics exert stronger effects in this model than the share of European minorities, but using different indicators than the 'White Other' census category (which includes West Europeans) produces different results. If, in place of 'white other increase,' we substitute a term for East European (born in post-2004 EU accession country) increase over 2001e11, this produces stronger 'white other' than 'visible minority' threat effects. There is also a 0.45 correlation between the 'visible minority' and 'white other' population shares at ward level, which complicates interpretation, though acceptable variance inflation tolerances (VIF<4) are not breached. All told, while these results might suggest UKIP voting is more sensitive to non-European than European minority levels and changes, this conclusion should be interpreted with caution.
The main point to take away is the pattern of disparate contextual effects from minority levels and changes on the dependent variable. Fig. 1 shows that a move from 0 to 50 percent minorities in ward reduces White British respondents' predicted probability of supporting UKIP from 0.02 to 0.005. Ethnic change predicts the opposite: a decline in minority share of 20 points during the decade of the 2000s is associated with a probability of voting UKIP of just 0.009. However, in wards with a 20-point increase in minority share over this period, this rises to 0.023.> 9

Self-selection tests
Our findings thus far show that higher shares of ethnic minorities in a White British individual's ward predict lower UKIP support. Contact theory would explain this as the result of positive contact between whites and ethnic minorities. However, threat theorists could legitimately demur, claiming that anti-immigration whites tend to disproportionately leave diverse areas leaving tolerant whites behind. This, not contact, best explains white tolerance in more diverse wards. Therefore I need to address the endogeneity, or self-selection, problem posed by 'white flight' and avoidance. As Abrajano and Hajnal (2015:151) comment, 'any further conclusions about neighbourhood context will have to wait until more rigorous testing can incorporate selection issues at lower levels of aggregation.' This paper addresses this concern.
To test for self-selection, I need to know whether UKIP-voting whites tend to disproportionately outmigrate from, or avoid, diverse areas. Our modelling strategy is to first test for white flight from, then to address white avoidance of, diverse areas by white domestic migrants. The UKHLS, as a longitudinal, large-sample survey containing questions on voting, permits us to do this in a way that is not yet possible with American data. The PSID contains no political questions and the GSS panel survey arguably lacks adequate sample size e for instance, there were just 40 white individuals who moved tract within their county in the 2008e10 GSS panel (Schachter, 2015). I attach census data for 1991, 2001 and 2011 assigned to a common ward geography.
Our first model asks whether UKIP-voting White British respondents are more likely to leave a ward with a large share of non-European minorities than non-UKIP voting White British respondents. The dependent variable in model 1 is a dummy coded 1 for a move out and 0 for remaining in, or moving within, a ward. 10 If threat theorists are correct, I would expect UKIP voters to be disproportionately represented among those leaving diverse areas, provided I control for confounding predictors of mobility such as age, income, marital status and education. I restrict the analysis to White British respondents.
The dependent variable in model 2, which is restricted to White British movers, is the difference in minority share between origin and destination ward. This measures how the share of minorities changes with a white individual's move. An individual who moves from a ward with a large share of minorities to a homogeneously white ward scores a strong negative (in theory up to À100% minorities) whereas a respondent who moves from a homogeneous to a diverse area would show an increase in ward minority share of up to þ100%. The dependent variable could in theory span À100 to þ100 though the actual range in the data is À79 to þ86. If threat theorists are correct, I would expect UKIP voters to be overrepresented among those in the high minus category (leaving minority-rich wards) while being underrepresented in the flow towards them. Model 1 considers white flight. Here the key parameter is the interaction between UKIP support and share of minorities in origin ward. The data show a small effect for minority share, suggesting that White British respondents are more likely to move from (rather than remain in) diverse wards as compared to white ones, but within the white flow, UKIP voters do not stand out. That is, whites who back UKIP and live in diverse wards are not significantly more likely to leave their ward than non-UKIP voting whites in similarly diverse wards, or UKIP-voting whites who live in lilywhite wards. This casts doubt on threat theorists' contention that self-selection of UKIP supporters out of diverse wards accounts for the relative tolerance, i.e. low UKIP voting, of whites in highminority contexts.
Model 2 examines white avoidance. The higher the share of minorities in a respondent's ward of origin, the bigger the drop in the share of minorities experienced as a result of a move. This is mainly an artefact of the supply of alternative wards: any person, white or otherwise, who lives in a high minority ward has very few higher-minority wards to choose from. This said, the data (not shown) reveal that whites leaving high-minority wards choose significantly whiter areas to move to than minorities originating from similar areas even when population density and deprivation are accounted for. But the main effect and interactions for UKIP support are not significant. In other words, while whites move to whiter wards than minorities e all else being equal, UKIP and non-UKIP supporting whites move to equally white wards. This supports H5.
We see this in Fig. 2, based on Table 2, where the confidence intervals for the red line for UKIP-voting whites and blue line for non-UKIP voting whites overlap. That is, among whites, with a range of individual and contextual predictors held at their mean values, both UKIP and non-UKIP voting respondents originating in wards with 46 percent minorities at time t -1 tend to move to wards which contain 13 points fewer minorities (i.e. are 33 percent minority) at time t. 11 Once again, I find no support for the view that the self-selection of UKIP supporters out of diverse wards explains the finding in Table 1 that higher local minority share is associated with significantly lower UKIP voting. This confirms H5 and suggests that contact rather than threat best explains our findings with regard to minority levels.
Immigration attitudes, much more than social background, predicts UKIP support (Evans and Mellon, 2016). I therefore replace UKIP support with attitudes to immigration as the dependent variable to probe for a similar disjuncture between the effect of ethnic levels and changes on white threat perceptions.

Dependent variable
The dependent variable for this study is the Citizenship Survey 11 Note that the typical drop in minority share with such a move is actually much larger. That is, if one does not control for independent variables such as deprivation or population density, the decline in minority share rises because diverse wards are also urban and poor while the suburban, wealthier destination wards which attract more movers tend to be considerably whiter (Catney and Simpson, 2010). question, 'Do you think the number of immigrants coming to Britain nowadays should be changed?' Answers follow a 5-category ordinal scale: 'increased a lot', 'increased a little', 'stay the same', 'decreased a little,' 'decreased a lot.' This variable is far from normally distributed, with a heavy slant toward reduction. Thus it has been recoded as two distinct dependent variables. The first is a binary reduce/do not reduce dummy variable which groups the roughly 81 percent of UK-born whites who desire a reduction into one category (1), and those favouring the same or more immigration (0) into a second. A second formulation isolates a dummy variable in which the approximately 60 percent of whites who desire that immigration be reduced 'a lot' are coded 1 and others 0. The analysis is restricted to white, UK-born residents. 12

Independent variables
I test a variety of individual-level parameters. Demographic and economic variables include age, marital status, sex, income, education, accommodation type (renter, owner, council tenant). Contextual parameters are drawn from the 2011 census except where noted. These include the proportion of the ward of residence comprised of ethnic minorities and the rate of minority increase in the ward since 2001. I consider the share of unemployed and population density in a ward, as well as minority share at Local Authority (LA) level.

Results
Table 3 presents a logistic regression of White UK-born immigration opinion (reduce vs. same/increase) on individual and contextual variables. Only intercepts (not slopes) for contextual variables are considered. As expected from the literature, I find that older respondents and those without formal educational qualifications are more opposed to immigration. Women do not differ from men in their attitudes. Critically for this paper, I find that the  level of minority share in a respondent's ward of residence predicts reduced opposition while the change in minority population between 2001 and 2011 is linked with greater opposition to immigration. Those in more deprived wards evince greater opposition. The pattern of contextual effects aligns with the models in Table 1 based on UKIP support in the UKHLS. 81 percent of White UK-born respondents favour reduction and 60 percent favour the 'reduce a lot' option. Yet a similar model emerges when I alter the dependent variable from reduce vs. same/increase to reduce a lot vs. reduce a little/same/increase. The black set of lines in Fig. 3 shows that the predicted probability of a White UK-born respondent favouring reduced immigration, with all other variables held at their means, falls from 0.85 in a ward with no minorities (denoted 'ML0') to 0.62 in a ward comprised of 50 percent minorities, denoted 'ML50'. Here it is worth noting that 80 percent of the 8850 wards in England and Wales are highly non-diverse: averaging just 6 percent minorities in 2011 while 41 percent of ethnic minorities live in little more than 400 wards, which average just 40 percent white (Kaufmann and Harris, 2014: 52). Widening confidence intervals suggest less robust inferences as the share of minorities rises toward 50 percent, but the pattern is clear and remains statistically significant, reinforcing the claims of contact theory.
Against this, the red line for minority change in ward shows that as we transition from the sample limit of À4, denoted 'MC-4' (i.e. 4 points fewer minorities in the respondent's ward in 2011 than in 2001) to the sample maximum of þ36 (a 36-point increase in minority share in a respondent's ward between 2001 and 2011), denoted 'MC36', this corresponds to a rise in the predicted probability of favouring reduced immigration from 0.77 to 0.90. The effect is less than for levels, and is also subject to wider error over much of its span, but is nevertheless statistically significant.
Repeating the analysis with 'reduce a lot' as the dependent variable (see Fig. 2) reveals a more pronounced pattern. For instance, the predicted probability of a White British respondent favouring reducing immigration by 'a lot' declines from 0.67 to 0.36 as one moves from zero to 50 percent minorities in the respondent's ward. On the other hand, the predicted probability of opposition rises from 0.53 to 0.76 comparing respondents in wards with the lowest and highest ethnic minority increase over 2001e11. This supports H1 and H2.
Throughout we see the divergent contextual effects of local ethnic levels -which predict a contact effect of reduced opposition to immigration; and local ethnic changes, which predict a threat response of heightened opposition. This said, the contact effects have a stronger effect size, which may be gleaned from the slope of the black compared to red lines in Figs. 3 and 4. Contact effects are also more robust, as indicated by the wider confidence intervals for the red lines.

The Local Authority context for UKIP voting in the 2014 European elections
Thus far, our analysis of anti-immigration sentiment and voting has concentrated on the ward, a unit averaging around 6500 individuals, as a measure of local context. I noted that higher minority levels are associated with lower, and faster minority changes with higher, anti-immigration sentiment and voting. What happens when we move up to the level of the Local Authority where population averages just over 200,000?
The British Election Study (BES) 2015 Combined Internet Panel Study permits us to examine the effect of Local Authority (LA) ethnic levels and changes on reported UKIP vote. The dependent variable is a dummy for reported UKIP vote in the May 2014 European elections (1 ¼ UKIP vote, 0 ¼ No UKIP vote, including nonvoters). I also probe immigration opinion based on a 7-point scale for the question 'Immigration undermines or enriches cultural life' from 1-undermines to 7-enriches. In order to minimise reverse causation, dependent variables are based on responses in wave 2 while independent variables are drawn from what the respondents reported in wave 1. Only White British respondents are included, which reduces sample size somewhat to 21,660. There are 5348 UKIP voters in the White British sample, representing 24.7 percent of White British voters, and thus somewhat of an undercount compared to the 26.6 percent obtained by UKIP in the election. 13 The analysis comprises three models, shown in Table 4. The first is an Ordinary Least-Squares (OLS) linear regression of antiimmigration opinion based on the 'undermines cultural life' question. This has also been run as an ordered logit and the coefficients are similar. Older, poorer, less educated and female respondents are more anti-immigration. The gender finding seems counterintuitive since the Citizenship Survey models in Table 3 show that gender is not significantly associated with immigration attitudes (Kaufmann and Harris, 2015). More germane to this study, however, is that the familiar effects of levels and changes assert themselves at the LA level. A 1 point increase in LA district minority share in 2001 is associated with a 0.12 of a point decrease in anti-immigration sentiment on a 7-point Likert scale. Conversely, a 1 point rise in the rate of minority increase in a respondent's district corresponds to a 0.15 point increase on the anti-immigration scale.
Moving to measures of the UKIP vote in the European elections, we find that younger voters and women are significantly less likely to have voted UKIP than men, reflecting established scholarship (Ford and Goodwin, 2014). Poorer and less educated voters are, by contrast, more likely to report voting UKIP. The share of minorities in 2001 in a respondent's LA is at borderline statistical significance, and signed in a negative direction, as in the anti-immigration model. Similarly, the coefficient for minority change shows a significant positive value. The final model includes the antiimmigration variable among the explananda on the right-hand side of the equation. This reduces the statistical power of minority levels and changes to some extent, though minority change remains important. This suggests that part, but not all, of the effect of ethnic context at the LA level on UKIP voting operates via immigration attitudes. Now consider Fig. 5. In terms of marginal effects, based on the 'UKIP1' model in Table 4, a move from no ethnic change in a respondent's Local Authority to a 28-point increase, the limits of this sample, corresponds to an increase in the predicted probability of reporting a UKIP vote from 0.20 to 0.32 with other variables held at their means. Minority levels in a respondent's LA just fail to reach significance at the 0.05 level in the 'UKIP1' model. The predicted probability of a UKIP vote varies by four points such that the marginal probability of a UKIP vote is 0.23 in the least (1%) and 0.19 in the most (46%) minority-rich Local Authority when other variables are at their mean values.
A shortcoming of the above analysis is that I cannot be sure of the extent to which ethno-contextual dynamics at lower  geographies such as wards, which comprise Local Authorities, are driving the results or whether higher-level contextual effects are key. In Table 1, using Understanding Society data, LA minority levels and changes were significant and showed effects broadly similar in size to Table 4. However these washed out when ward levels and changes were added to the model. Previous work would also suggest that both levels and changes at higher geographies (i.e. LA) may be associated with white opposition to immigration -as noted by Abrajano and Hajnal for the US case (2015: 132). Until the BES can negotiate access to ward census data for BES respondents, it is not be possible to ascertain whether higher or lower-level dynamics are central.

Toward ethnic accommodation?
Today's ethnic changes feed into tomorrow's ethnic levels. This presents a paradox e how do threat effects associated with change transform into contact effects associated with levels? No paper I am aware of addresses this question. The answer appears to be that yesterday's demographic changes fade through habituation, legitimation and contact while yesterday's ethnic levels become increasingly potent predictors of lower threat levels over time.
As Table 5 shows, minority levels gain in predictive power as one goes back in time, confirming H4. That is, the coefficient for 1991 minority level exceeds that for 2001 or 2011 minority level in both the UKHLS and BES. Moreover, in the BES, minority share only attains significance in 1991. One could argue that this is because minorities in wards with a long experience of minority settlement are viewed as 'established' which legitimates their presence and hence reduces white residents' sense of dispossession and, by extension, lowers support for UKIP. Contact may not therefore be the only mechanism at work translating minority-rich contexts into more liberal attitudes to immigration. The figure likewise reveals that only 2001e2011 ethnic change significantly predicts UKIP support. 1991e2001 change in the same wards, though displaying similar or larger coefficients, is not significant. 14 H3 is thereby also confirmed.
Taken at once, these results indicate a pattern of shock, followed by habituation. The freshest ethnic changes have the most dramatic threat effects on white opinion. Conversely, the older the 'tradition' of high levels of diversity, the stronger the impact of contact effects. This suggests two things: first, that part of diversity's mollifying effect may operate through legitimation over decades rather than mere contact with immigrant groups (though it may be that interethnic contact intensifies over time); and second, that periods of rapid ethnic change correlate with heightened concern over immigration and far right voting because threat effects are at their maximum while contact effects have yet to build momentum.

Local election results
As a final robustness check I examine local level election data. Actual election results contain a far larger sample of UKIP voters (hundreds of thousands) than surveys, and are uncontaminated by social desirability and other response bias. I examine a set of local election results compiled during 2010e12 by the Elections Centre at the University of Plymouth, around the time of the 2011 British census. Local elections only take place across a subset of districts in Britain each year thus a full set of results can only be accumulated over several years. 2010, 2011 and 2012 election results by ward are linked to a dataset containing 2011 ward-level census data. UKIP's performance is also assessed against that of a more radical antiimmigration party, the British National Party (BNP). The BNP enjoyed greater popularity than UKIP in local elections until after 2009, when it fell prey to leadership splits and scandals (Harris, 2012).
I only examine wards in which the UKIP and BNP stood candidates. Table 6 shows that UKIP ran in 945 wards over this period and the BNP in 704. Results in the first model in Table 4 show that the BNP vote conforms to the previously-noted pattern whereby high minority levels in a ward predict lower anti-immigration sentiment and voting while faster minority change is associated with elevated anti-immigration sentiment and voting. Indeed, this has been the consistent finding of previous work on the BNP (Bowyer, 2008;Biggs and Knauss, 2012;Ford and Goodwin, 2010;Harris, 2012). Replicating the analysis with UKIP rather than BNP vote as the dependent variable, I find that the first, but not the second, relationship holds. Results are similar when I divide the dependent variable by the White British share of the population to exclude minority voters. Of course, aggregate results are susceptible to the ecological fallacy (Robinson, 1950). Yet with a properly specified model, these concerns can be minimised even as model fit statistics may be inflated because both voting and ethnicity are geographically concentrated (Knoke, 1974).
Results are graphically reflected in Figs. 6 and 7. The x-axis combines minority share and minority change (a combination of low share and fast change is predicted to maximise threat effects), which is plotted against BNP or UKIP vote on the vertical axis. Notice that the slope for BNP vote share is considerably steeper than for UKIP vote share, reflecting the fact that ethnic change is only significantly associated with elevated populist right voting in the BNP case. This may be because UKIP declined to compete with the BNP in strongly anti-immigration wards in the 2010e12 period: 14 Naturally such analyses are complicated by multicollinearity: the correlation between 2001 and 2011 change and 2011 levels is 0.64. Nonetheless, the analyses in Table 5 show acceptable variance inflation statistics (VIF<4) when run as OLS regressions.
post-2012 it did especially well in Barking. This is an avenue for further research. A clutch of wards in Barking and Dagenham, a Local Authority whose White British population declined from 81 to 49 percent during 2001e11, show very high BNP support. Results are robust to excluding Barking from the analysis, but it is worth noting this as a paradigm case of the effects of rapid ethnic change. The BNP shocked observers by winning 12 of 51 council seats in 2006, leading to a high-profile 2010 campaign, led by Margaret Hodge of Labour, which mobilised newcomers and former nonvoters to turn out and vote Labour, resulting in the BNP losing all its councillors. This despite the BNP doubling its vote share between 2006 and 2010. 15 In fieldwork conducted in 2012 in Barking and Dagenham, many white respondents cited ethnic change as a motor of BNP support. 'I voted BNP,' reported a respondent named Eleanor, 'I can't help it. They call them Nazis. But they're not. They're Britain for Britain. Labour sent [immigrants] all down here and [Margaret Hodge] won't tell me where they come from. I think they fiddled the votes, so that the BNP did not get one candidate in.' 16 While UKIP also did well in several Barking wards, and by 2014 had equalled the BNP's 2006 vote share in the district, its performance lagged behind the BNP during 2010e12. This does not appear to be due to competition in the 256 wards in the sample in which the two parties competed because results are similar when these are excluded. Therefore an interpretation of the difference between the BNP and UKIP in this period is that the pattern is affected by each party's social profile. In the UKHLS, the median White British respondent is 49 and lives in a ward that is 11.7 percent minority. By contrast, the typical BNP supporter is 43 and resides in a ward with 12.5 percent minorities while the median UKIP supporter is 57 and inhabits a ward with just 10.1 percent minorities. BNP voters are younger and live in significantly more diverse places than the average White Briton. UKIP voters are older and reside in more homogeneous wards than average. Since ethnic change in a ward is highly correlated with its existing level of ethnic diversity, BNP voters were more likely to reside in wards experiencing ethnic change than UKIP voters at this time.

National-level dynamics
What does the foregoing mean for the wider question of how nations will respond to rising immigrant diversity? In entire nations, does ethnic change elevate threat while higher established levels of minorities calm it? In the studies I have examined which use country as the geographic unit, 22 of 27 tests in 14 studies (13 of 14 in 6 studies reporting significant effects) find a threat effect from country-level increases in immigrants or minorities (see Appendix III and IV for details). By contrast, when restricted to 26 studies reporting significant results, country minority levels are associated with threat effects in 24, and contact effects in 13, tests (see Appendix V). Though higher minority levels predict threat in 65 percent (24 of 37) of significant tests at country level, this is well below the 92 percent (12 of 13) of tests reporting a threat-enhancing effect for minority change. These patterns are indicative. Given the results obtained in this paper, I would expect that if these works had incorporated a test for both levels and changes, the balance of tests would move toward an association between higher diversity levels and threat reduction.
Consider the British case, where a significant increase in immigration over the period 1997e2001, from roughly 50,000 to 250,000 per annum, took place. Fig. 8, drawn from Duffy and Frere-Smith (2014) compares actual net migration with the salience of immigration and race issues in the Ipsos-Mori issues index since 1992. The polynomial curves of the two lines exhibit a 0.78 correlation.
Since 2002, in the presence of higher annual inflows, the British public has ranked immigration as its first or second most important issue. A similar relationship may be observed when plotting net  migration (a measure of immigration) against results of the quarterly MP's Survey which asks Members of Parliament to list their constituents' leading concerns. Here again, the proportion mentioning asylum/immigration/refugees as their most important issue broadly tracks net migration figures from the late 1980s to January 2013. Likewise, since the high-immigration period beginning in the early 2000s, the proportion of MPs mentioning immigration-related issues as the primary concern expressed in their constituency mailbag has averaged around 60 percent (Duffy and Frere-Smith, 2014: 8e9). The relationship between arrivals and public opinion is somewhat of a lacuna in current research, which has typically focused on the downstream effect of coverage on opinion (i.e. Dunaway et al., 2010;Hopkins, 2011;Abrajano and Hajnal, 2015;ch. 5). Fig. 8 shows a marked correlation between national-level inflows and public concern over immigration. Nevertheless, this is not an unmediated response: Duffy et al. report that net migration correlates with negative news coverage in Lexis-Nexis. 17 Thus, the correlation could be viewed as multi-step, with rising immigration propelling increased media coverage and politicisation, interacting with lived experience of ethnic change to produce greater opposition to immigration. Future cross-national research is needed to test linkages between actual immigration numbers, media coverage and public opinion. My earlier results regarding local change may hold lessons with respect to national patterns. Namely, that if the pace of change slows and people habituate to higher levels, concern may fade, as in the post-Powell era. Note that this period habituation is a somewhat different argument from that which posits cohort attitude change producing a steady liberalisation in attitudes -though this too may involve younger cohorts becoming habituated to a 'new normal' of higher diversity (Ford, 2008;Putnam, 2007).

Discussion
The rise of the UK Independence Party (UKIP) has been the most disruptive development in British party politics for a generation, arguably helping mobilise the Brexit vote. This research claims that ethnic change stimulates higher UKIP support while higher levels of established minorities predict lower UKIP support. A similar pattern characterises immigration attitudes in Britain. These findings echo those from the wider ethno-contextual effects literature. Yet whereas some previous work (i.e. Newman, 2013) notes the greater importance of changes compared to levels in predicting threat effects, none I am aware of posit cross-cutting effects for levels and changes. Nor do they show as explicitly that such findings are not an artefact of self-selection: white UKIP voters in minority-dense wards are no more likely to leave them, or to move to whiter wards, than their white liberal neighbours.
This research also breaks new ground by suggesting that the threat effects of demographic change fade over time, possibly because local white residents become accustomed to immigrant minorities, are more likely to have positive contact with them, or come to perceive immigrant groups as having a legitimate place in society. This is important because while levels and changes exert disparate effects on white opinion, yesterday's immigration flows constitute today's immigrant stock. These results suggest that if the rate of ethnic change declines in the West in the future, opposition to immigration and support for the populist right could similarly abate.