Ethnic Spatial Segregation And Tobacco Consumption A Multilevel Repeated

ethnic spatial segregation and tobacco consumption: a multilevel repeated cross-sectional analysis of smoking prevalence in urban new zealand

Ethnic spatial segregation and tobacco consumption: a multilevel
repeated cross-sectional analysis of smoking prevalence in urban New
Zealand, 1981-1996
Graham Moon (1)
Ross Barnett (2)
Jamie Pearce (3)
1.
Corresponding Author: Centre for Geographical Health, University
of Southampton, Highfield, SOUTHAMPTON SO17 1BJ, UK.
[email protected]
2.
GeoHealth Laboratory, University of Canterbury, Christchurch, New
Zealand
3.
Institute of Geography, School of Geosciences, University of
Edinburgh, Drummond Street, EDINBURGH EH8 9XP, Scotland.
Abstract
========
The growing literature on the role of ethnic segregation in
understanding spatial inequalities in mortality and morbidity has not
yet been extended to the study of health-related behaviours. We
address this gap in knowledge through an examination of the geography
of smoking prevalence in New Zealand using a multilevel repeated
cross-sectional analysis of smoking prevalences in 1981 and 1996 as
revealed in the New Zealand census. Smoking prevalences are explored
for fourteen age and sex groups nested in 1110 census area units.
These in turn are nested in 40 primary and secondary urban areas. We
consider different measures of segregation and focus in detail on the
relationship between smoking and Māori ethnic isolation. We examine
the interplay between deprivation and segregation addressing questions
concerning the impact of changing segregation on changes in smoking
behaviour. We hypothesise that more highly segregated populations
suffer more psychosocial stress so may smoke more. Results reveal the
changing dynamics of smoking prevalence over time and challenge
initial assumptions that spatial ethnic segregation should relate to
smoking prevalence.
Introduction
============
Research on ethnic spatial segregation has a lengthy history. Rooted
in the Chicago School, there has been a resurgence of interest since
the 1980s, notably in measurement approaches such as dissimilarity and
isolation indices (Massey and Denton, 1987, 1988; Peach, 1996;
Phillips, 1998). Recently a growing, but equivocal, literature has
begun to examine the relationship between ethnic spatial segregation
and health outcomes (Acevedo-Garcia, 2000; Acevedo-Garcia and
Lochner, 2003; Williams and Collins, 2001). The present paper seeks
to develop and extend the latter body of research in three ways.
First, we shift the empirical subject of study from health outcomes to
health-related behaviour. While previous research has tended to focus
on mortality, particularly child mortality, this paper will consider
what is arguably the most significant health-related behaviour:
tobacco consumption. This shift in focus is conceptually important as
spatial and social variations in health-related behaviour are
implicated in a wide range of health inequalities. Second, past work
has tended to focus on the health impact of the often stark ethnic
segregation found in the USA. In contrast, the present paper considers
the health impact of segregation in a society where, although ethnic
disadvantage is omnipresent, there has been greater population mixing.
Third, we introduce a temporal element to the study of the
relationship between ethnic spatial segregation and health. Previous
work has generally involved single cross-sectional snapshot insights.
Both smoking behaviour and ethnic spatial segregation have changed
significantly over recent years; the paper seeks to uncover the ways
in which these changes inter-relate.
The research questions that motivate this paper concern the nature,
extent and direction of the relationship between ethnic spatial
segregation and smoking within an ethnically plural society exhibiting
moderate segregation. We examine how these relationships have altered
over time and varied over space, focusing on a period of major
economic and social change during which socio-economic inequalities
known to impact on smoking were enhanced. The paper has four
substantive sections. First, we examine the literature on health and
ethnic spatial segregation. Attention then turns to the empirical
setting for the study: Aotearoa/New Zealand. Separate examinations are
presented for smoking and for segregation. The third substantive
section sets out our data and analytical approach. Finally, we report
the results of our analysis and discuss their implications.
Ethnic Spatial Segregation and Health
=====================================
While a large body of research has sought to understand health
variations by incorporating some measure of ethnic status as a
predictor variable (Nazroo, 2003), our concern, in this paper is
with community-level measures of ethnic spatial segregation. Five
separate constructs have been associated with this theme, each
identifying different aspects of spatial segregation (Iceland et al.,
2002; Massey and Denton, 1988). (Un)evenness measures consider the
distribution of majority and minority ethnic groups within sub-areas
relative to the distribution in larger areas. Exposure or isolation
measures focus on the likelihood of neighbourhood-based contact
between ethnic groups. By far the majority of studies of health and
ethnic spatial segregation have been concerned with evenness or
isolation. The remaining constructs are concentration, referring to
the share of physical space occupied by majority and minority groups,
centralisation and clustering.
In terms of subject matter, most past studies have linked measures of
ethnic spatial segregation to infant health in the USA. Higher levels
of black-white segregation have been associated with infant mortality
in both black (Bird, 1995; Guest et al., 1998; LaVeist, 1993) and
white populations (Polednak, 1991, 1993). Similar conclusions have
been drawn concerning ethnic spatial segregation and low birth weight
(Ellen, 2000). Moving beyond infant health, there has also been a
limited amount of past research on adult mortality and ethnic spatial
segregation. This work has linked higher levels of segregation with
raised all-cause mortality, cardiovascular mortality and cancer
mortality among both black and white US residents (Collins and
Williams, 1999; Cooper et al., 2001; Hart et al., 1998). Throughout
these studies, there is general confirmation that the relationship
with ethnic spatial segregation persists after controlling for
socio-economic confounding but is perhaps more marked for men, and
segregation impacts outcomes for both minority and majority groups.
What the above studies have in common, in addition to their US
setting, is that they are single-level designs. Most are ecological
studies, examining area-level segregation measures in relation to
area-level outcome measures. A few work with data on individuals.
Studies are now increasingly using multilevel designs, recognising
that health outcomes may reflect, simultaneously, both individual
factors (age, sex etc.) and areal influences such as spatial ethnic
segregation. A multilevel study of black-white disparities in
self-rated health (Subramanian et al., 2005) used data on some
50,000 individuals nested in US metropolitan areas. The impact of
individual factors such as age, sex, income and marital status was
considered in relation to area level segregation measures. Subsequent
work has considered ethnic differences in body mass index (Chang,
2006) and birth outcomes (Bell et al., 2006; Grady, 2006; Wong,
2002). Grady suggests that segregation at an area level can
exacerbate the effect of area level poverty.
As yet, very few studies have examined smoking in relation to ethnic
spatial segregation measures. Among the few focussed papers is a small
individual-level study in California, USA of 520 black adults in ten
census tracts (Landrine and Klonoff, 2000). This research used a
non-standard 28-point individual segregation score and concluded that
segregation was associated with a significantly higher rate of smoking
and was not an artefact of socioeconomic status. A second
demographically-restricted study (Ellen, 2000) concluded that
tobacco use by black women during pregnancy was related to segregation
in the form of ethnic centralisation in urban areas, but not to
dissimilarity (a measure of unevenness). Smoking in pregnancy was also
examined in two further multilevel studies: it is relatively more
common in the US at high or low levels of segregation rather than
moderate segregation (Bell et al., 2007) while in Sweden the reverse
appears to be the case (Sellstrom et al., 2008) .
Underpinning the studies reviewed above are psychosocial hypotheses
about health inequalities (Brunner, 1997; Kawachi et al., 2002;
Siegrist and Marmot, 2004). On the one hand, segregated groups may be
expected to experience greater stress and hence worse health. On the
other hand, positive norm reinforcement may be easier when segregation
is high with people of similar ethnic status living together (Stuber
et al., 2008). The balance of evidence is overwhelmingly suggestive
of a negative impact of but some research has hinted at the more
positive associations. Thus black women living in areas that are
mostly black, particularly if those areas are relatively affluent, are
less likely to have under-weight or premature babies (Pickett et al.,
2005; Roberts, 1997). The potential pathways by which this divergent
perspective may eventuate are perhaps best understood by recognising
the relationship between ethnic spatial segregation and
discrimination. In highly segregated areas, a common experience of
poverty allied with ethnic homogeneity may indicate shared levels of
disadvantage that reduce the psychosocial stresses evident in
neighbourhoods where more advantaged others of different ethnicity are
more common. The causal pathway of discrimination – stress – morbidity
may be clearer in areas with a greater ethnic mix. Of importance,
therefore, is the need to untangle confounding by socio-economic
status, both at an individual and a neighbourhood level.
In summary, current work on health in relation to measures of ethnic
spatial segregation has had a dominant focus on mortality,
particularly infant mortality, and an overwhelming concentration on
the US. It has drawn mixed conclusions concerning the positive or
negative health consequences of segregation and the possible
confounding effects of socio-economic status. Some work considers
segregation in relation to the general population, other work focusses
on specific populations. Though we have not discussed the point in
detail, existing research has also used several different measures of
ethnic spatial segregation, often with limited theoretical
justification.
On the basis of the above assessment, we identify a clear case for
research on smoking and segregation testing associations with
different measures of segregation and controlling for socio-economic
confounding. We also see a clear case for extending research to
settings other than the US. It is further hypothesised that an
association between segregation and smoking should be generally
evident within a population. Both majority and minority groups within
segregated populations experience the psychosocial stresses that have
been associated with a greater prevalence of smoking. Areas with
higher segregation should thus have raised levels of smoking. We see
an exploration this generalised impact of segregation as a necessary
precursor to more detailed work on outcomes specific to particular
ethnic populations.
Smoking and segregation in Aotearoa/New Zealand
===============================================
Levels of smoking in New Zealand accord largely with expectations for
countries in the late phase of the smoking transition1 (Lopez et al.,
1994; Shaw et al., 2005). Data from the 2006 New Zealand Tobacco Use
survey suggest that some 23.5% of the population smoke: a relatively
low figure on an international scale (Ministry of Health, 2006).
Smoking has been declining for over thirty years and prevalences are
now roughly balanced between the sexes and year-on-year rate
reductions are levelling off. Smoking by women is however not now
currently declining at the same rate as male smoking. Smoking among
young people is similarly problematic in that rates are highest
amongst people aged under 25.
Notwithstanding reducing prevalences, smoking in New Zealand is not
only patterned by age and gender, it also exhibits strong associations
with deprivation and disadvantage. These associations are evident at
both an area-level and between individuals. Deprived localities,
identified using the New Zealand Deprivation Index (NZDep), a
composite measure of area indicators of disadvantage, return
significantly higher levels of smoking (Crampton et al., 2000; Tobias
and Cheung, 2001). Smoking prevalence rates are approximately twice
as high in the most deprived quintile of neighbourhoods compared to
the least deprived quintile (Ministry of Health, 2004). At an
individual level, smoking in New Zealand has been found to be
associated with being single or divorced, manual work, low income,
poor education and overcrowding (Borman et al., 1999; Hill et al.,
2003; Howden-Chapman and Tobias, 2000).
These factors separately and collectively intersect with patterns of
ethnic disadvantage in New Zealand and it is not surprising that
smoking also exhibits marked differentiation by ethnic group. The
current smoking prevalence for New Zealanders of European origin
(henceforth Europeans), is estimated at 20.6 (Ministry of Health,
2006). The figure for Māori, who comprise some 14% of the New Zealand
population, is 39.6% and the smoking prevalence among Māori women in
2006 was 42.9%. In part this excess tobacco consumption may reflect
the relatively more youthful demographic profile of Māori but it also
points to a powerful interaction between ethnicity and deprivation
that underpins tobacco consumption in New Zealand. This interaction,
in turn, links to a marked geography to smoking in New Zealand, with
higher rates concentrated in more deprived areas with higher Māori
populations (Barnett et al., 2004).
There are further temporal dimensions to this situation. The disparity
in smoking prevalence between Māori and Europeans was not always
present. Its emergence parallels the development of socio-economic
differentials in smoking prevalence in the period since the 1960s and
a general widening ethnic gap in health inequalities in New Zealand
(Robson, 2003). Reductions in tobacco consumption over time have
been much more marked among Europeans, although rates have been
declining amongst both Māori and Europeans (Laugesen and Clements,
1998). The widening differential in smoking status between Māori and
non-Māori/non-Pacific peoples is a significant and growing contributor
to ethnic inequalities in health in New Zealand (Blakely et al.,
2006). As a consequence of this underlying temporal dynamic it is
clear that the achievement of further reductions in smoking prevalence
in New Zealand relies heavily on increasing Māori cessation rates and
narrowing the smoking gap between ethnic groups.
Temporal and spatial matters also characterise ethnic segregation in
New Zealand. In 1961 ethnic segregation was, in international terms,
identifiable but relatively low (Christopher, 1992). This has
continued to be the case in the intervening years and a distinction
has arisen between urban and more rural areas. Urban Māori segregation
in 2001 was greatest in larger urban areas with sizeable Māori
populations (Johnston et al., 2005a). Auckland, New Zealand’s
largest city, was an exception where the sharing of residential space
with Pacific peoples reduced Māori levels of segregation. Changes in
segregation over recent years have been closely linked to the growth
of populations of non-Māori non-Pacific origin and Asian origin. In
contrast to segregation in urban areas, Māori segregation in rural
communities and small towns has long been, and remains, generally
higher in absolute terms, than in larger urban areas (Johnston et
al., 2003; Poulsen et al., 2000). In both urban and rural areas,
there is a strong overlap between segregation and socio-economic
disadvantage. In neither urban nor rural areas however, does ethnic
spatial segregation generally approach the levels of segregation found
in the USA (Poulsen et al., 2001, 2002). Johnston et al. (2003)
note that, in 2001, it was the European population that exhibited the
greatest ethnic spatial segregation. Just 8% of Māori lived in
meshblocks (small census areas) where Māori provided over 60% of the
population. These Māori concentrations were heavily located in North
Island.
Overall, the empirical evidence suggests that, while segregation is
present, New Zealand is a relatively mixed society in terms of
ethnicity (Poulsen et al., 2000). How then might smoking and ethnic
spatial segregation be expected to inter-relate in the specific New
Zealand context? On a general level it is evident that New Zealand
provides a test of the extent to which a health issue is associated
with segregation in a less segregated setting than that prevailing in
the USA. More specifically, limited past research offers some
intriguing insights into possible findings. Barnett (2000) suggested
that socio-economic segregation exacerbated the impact of deprivation
on smoking. It would be reasonable to assume that this would include
ethnic segregation. A multilevel study using 1996 census data
concluded that the size of the Māori population relative to that of
Pakeha was more important for smoking levels than the absolute size of
the Māori population (Moon and Barnett, 2003). This study also
suggested that small-area Māori smoking prevalences were highest when
the small area in question was located within a larger area dominated
by Māori. However, isolated Māori populations in areas with otherwise
low Māori populations also had raised smoking prevalences. Finally,
Barnett et al. (2005) noted that increasing ethnic inequalities in
deprivation between 1981 and 1996 were associated with higher levels
of smoking, confirming the importance of simultaneous consideration of
deprivation when analysing the relationship of ethnic segregation to
smoking and pointing to a need for a temporal dimension in future
research.
Research Strategy
=================
The case for a multilevel approach to research on ethnic spatial
segregation and health has been well-made (Acevedo-Garcia et al.,
2003; Osypuk et al., 2006; Subramanian et al., 2005) and include,
within limitations, an ability to address aspects of the ecological
fallacy. The conflicting conclusions of past research may, in part,
reflect a general failure to adopt appropriate multilevel analytical
strategies. In the present paper our outcome of interest, smoking
behaviour, can be conceptualised as a consequence of individual and
areal factors. A multilevel research strategy allows us to separate
out the independent impacts of individual and areal level influences.
Moreover, multilevel analysis enables further consideration of more
complex relationships between individual and areal factors, for
example the extent to which variables at different levels interact to
influence an outcome variable (Duncan et al., 1996, 1999).
The data for this study were drawn from the New Zealand census. The
inclusion of a question on smoking in selected New Zealand censuses is
a major advantage to research on smoking where more usually reliance
is placed on sample survey data. The census provides information on
smoking (and ex-smoking) that is as complete as possible. The census
smoking question focusses on intentional smoking of one or more
cigarettes per day by individuals aged fifteen or over. We chose
deliberately to work with the censuses for 1981 and 1996. This time
period was purposefully selected for comparability with previous work
on smoking in New Zealand (Barnett et al., 2005) and because
1981-1996 was a period of major socio-economic change in the country.
The economic reforms of this period (Le Heron and Pawson, 1999)
brought increasing ethnic segregation as well as ethnic socio-economic
inequalities, providing optimal conditions for the psychosocial
stresses that have been associated with smoking behaviour. Work on
health inequalities has demonstrated how disparities increased
markedly during this period and then hit a plateau (Pearce and
Dorling, 2006) while political commentators have noted the impact on
inequality and the later retreat from the associated policy
perpectives (Mitchell, 2005).
The measures included in the study were compiled at three spatial
levels, reflecting the desired multilevel design. Individual level
data (level one) were assembled from a three-way cross-tabulation of
smoking status (smoker/non-smoker) by age (seven classes) by sex for
each census area unit (CAU – areas containing 3-5,000 people). The
cell structure of these tables identifies the age-sex smoking status
of each individual census respondent; though the level represents
individuals, it might more appropriately be referred to as cell-level.
We focussed on age-sex smoking prevalences, without further
disaggregation to identify ethnic-specific smoking, for both
theoretical and pragmatic reasons. As noted above, existing research
is equivocal on the impact of segregation but there are theoretical
grounds for expecting a generalised impact on smoking. A necessary
first step in exploring the impact of segregation on smoking in New
Zealand is thus to identify the impact of segregation on overall
levels of smoking. More pragmatically however, data disclosure risks
meant that the four-way CAU-level age-sex-smoking-ethnicity
cross-tabulations that would have enabled us to model ethnic-specific
smoking were not released to us. We were thus constrained to work with
just age and sex and to focus on the general impact of segregation;
further individual data were unavailable.
At level two, socio-economic data were collated at the level of the
CAU. A standard suite of indictors were used identifying different
forms of socio-economic deprivation: percentages of the CAU population
who were unemployed, resident in social housing, drawing domestic
purposes benefit (DPB), and without educational qualifications. This
set of data was chosen after a process of testing for
multicollinearity in which a fifth indicator, low income, was dropped.
The third level in our data set was the ‘urban area’ (UA). Ethnic
spatial segregation was measured at this level. We limited our study
to urban areas as rural areas are very sparsely populated in New
Zealand with consequences for data stability. Urban areas are
well-defined, commonly-understood entities with a reasonable degree of
congruence with housing markets and other structures that underpin
ethnic segregation2. A majority of the US studies reviewed above
considered ethnic spatial segregation at the US equivalent of the UA,
including the more robust multilevel studies. We included both main
(30,000+ population) and secondary urban areas (population
10,000-29,999) and incorporated an indicator variable to capture this
difference. Common CAU and UA boundaries and data definitions for 1981
and 1996 allowed the census year to form a fourth ‘level’ in the data
set. The complete data set is summarised in Table One.
Table 1: Multilevel Data Structure
1981
1996
Variables
Individuals (Cells)
15,176
15,589
Smoker/Non-Smoker
Age (7 bands)
Sex (Male/Female)
CAUs
1,101
1,124
DPB
Unemployed
No Qualifications
Social Housing
UAs
39
39
Main/Secondary UA
Dissimilarity
Isolation
Thiel’s ‘H’
NB Cells do not equal CAUs x 14 as cells with zero population totals
are omitted
Three standard measures of ethnic spatial segregation were calculated
at the UA level following established formulae3. The dissimilarity
index is that most commonly employed in general studies of ethnic
segregation. It is a measure of ‘evenness’ (Massey and Denton, 1988)
and ranges from 0 to 1 (complete segregation) We calculated Māori –
European dissimilarity. Our second index was an exposure measure:
Māori isolation. Higher values, up to unity, again indicate greater
segregation and capture the extent to which a minority group is
exposed only to its own members It has been extensively used in
previous studies of ethnic spatial segregation and health, and is the
measure of choice for such studies (Acevedo-Garcia and Lochner, 2003;
Subramanian et al., 2005) as isolated groups are more likely to
experience concentrated disadvantage with health outcomes. We used the
adjusted form of the isolation index (Jargowsky, 1996; Reardon, 2006;
White, 1986). This enabled us to take account of the relative size of
the Māori population (Johnston et al., 2005b). The third calculated
measure was Thiel’s ‘H’, an entropy index that describes the diversity
of an urban area (Iceland, 2004). ‘H’ has the same properties as the
other two indices but is multivariate rather than binary: it measures
how evenly all groups are distributed across an urban area rather than
focusing solely on a contrast between two groups. To this end, in the
New Zealand context, it allows recognition of the presence of ethnic
groups other than Māori and Europeans.
The conceptual framework for our study is summarised in Figure One. We
assume, on the basis of previous evidence (Duncan et al., 1999),
that the probability of being a smoker is primarily a reflection of a
person’s age and sex. At issue are the extent and direction of
ecological or area effects, and specifically the relative importance
of deprivation and segregation (Ross et al., 2001). To this end, our
interest centres on the direction of the causal pathway and the extent
of confounding between segregation and deprivation (Goldman, 2001).
Does segregation influence smoking through deprivation (continuous
arrows pathway) or does deprivation influence smoking though
segregation (dashed arrows pathway)? Indeed does either construct
influence smoking once individual factors are taken into account?
Figure 1: Conceptual Model

Following an initial exploration of our data on smoking and
segregation, we use an ecological analysis at the UA level to select a
measure of ethnic spatial segregation. We then address the above
questions though the development of four multilevel models. Model A
considers only the impact of demographic level one factors on smoking.
Model B controls for individual level main effects and brings in
segregation. Model C provides a similar analysis but for deprivation.
Model D examines both the individual main effects and, simultaneously,
the two higher level constructs of deprivation and segregation.
Comparisons between the models allows us to draw tentative conclusions
regarding causal pathways.
All models were estimated using MLwiN 2.02 (Rasbash et al., 2005).
To avoid working with data on the smoking habits of each individual
New Zealand resident, we used the binomial proportional modelling
strategy (Moon and Barnett, 2003; Subramanian et al., 2001). In this
way the cell structure outlined in Table One enabled us to model, as
an outcome variable, the proportion (of individual) smokers in each
age-sex cell. Indicator variables were used to identify the age and
sex status of each cell. Ecological, higher-level predictors were
centred on their 1981 value to aid the measurement of change between
1981 and 1996. We modelled using a logit link and second order
Penalized Quasi-likelihood (PQL) estimation (Goldstein and Rasbash,
1996). Random slopes for segregation and deprivation were
investigated and rejected as they did not enhance the models. Our
final models were random intercept models and were reviewed for
extra-binomial variation and recalibrated using MCMC procedures.
Results and Discussion
======================
Table 2 provides baseline information on the age-sex prevalences of
smoking for New Zealand as a whole for 1981 and 1996. The data conform
to established understandings of the evolution of smoking behaviour
during the period in question (Borman et al., 1999; Easton, 1995).
In 1981 men were considerably more likely to smoke than women at older
ages; in 1996 this sex gap had reduced. In both 1981 and 1996 women
smoked more than men in the youngest age group (15-25) and this sex
gap widened marginally over time. Overall the picture is the standard
western one of the feminisation of smoking with men reducing their
prevalence more sharply from higher levels and younger women
constituting something of a problem group, albeit one that needs to be
understood in the context of the gendered cultural significance of
smoking (Amos, 2001; Denscombe, 2001).
Table 2: Observed Smoking Prevalences (%)
Age
Men
Women
1981
1996
1981
1996
15-25
34.43
27.26
36.27
29.59
26-35
37.25
30.32
33.49
28.49
36-45
38.65
27.33
32.70
24.86
46-55
37.75
24.85
31.10
22.01
56-65
32.86
20.27
25.47
16.05
66-75
26.97
14.31
16.15
11.10
75+
19.11
9.11
8.32
6.29
The baseline UA geography of smoking in our two study years is set out
in Figure 2. The UAs are ordered in terms of their smoking prevalence
in 1981. The reduction in smoking between the two years is clearly
evident and present in all UAs. There is a very strong correlation
between the geographies of smoking in the two years (r=0.98). Smoking
in both years is also strongly related to the Māori percentage of the
population in each UA (r=0.72 (1981), r=0.75 (1996)) being
consistently higher in UAs in central and eastern North Island such as
Tokoroa, Rotorua and Taupo, where the Māori population is high.
Figure 2: Baseline Geography of Smoking, 1981 and 1996, Urban Areas

Smoking is also high in UAs in the greater Wellington area, such as
Porirua and Upper Hutt, and in Gore, a UA in South Island highlighted
by Moon and Barnett (2003) as having unusually high levels of
smoking given its location and ethnic characteristics. Greatest
reductions in smoking tend to have occurred in UAs where the scope for
reduction is highest: in UAs where smoking was high in 1981. There are
also notable reductions in Wellington and parts of Auckland,
suggesting an influence linked to the increasing connectedness of New
Zealand’s main cities during the 1980s to the growing smoking
cessation movement. Main or secondary urban status (denoted by a UA
with an asterisk) appears to have little effect.
The relationships between smoking and our three measures of ethnic
spatial segregation are explored at the UA level in Figures 3a-c. Each
measure is, as expected, positively correlated with smoking, offering
initial ecological support to the hypothesis that segregation affects
smoking in New Zealand. There is also variability within the observed
ranges of each index. On all measures, the overall relationship
between smoking and segregation is remarkably similar in 1981 and
1996. The isolation index has the strongest relationship to smoking
(r=0.49 (1981); r=0.46 (1996); both significant p=0.01). Outliers are
perhaps least marked on this graph and more clearly within the general
trend though a single high smoking outlier is evident in both years
(Tokoroa). The comparison correlations for Theil’s ‘H’ (0.38, 0.32)
and Māori –European dissimilarity (0.23, 0.32) were not statistically
significant. The distinction between the isolation index and the other
measures is also evident with regard to changes in segregation between
1981 and 1996. All areas saw increases on the isolation index (bar one
, which remained unchanged), with a mean rise in segregation of some
three percent. In contrast a miniscule overall rise in the Theil index
resulted from rises in multiway segregation in just 25% of UAs and no
change in a further third. Dissimilarity tended to reduce, though by
less than one percentage point with rises of up to 6% being evident in
parts of Greater Auckland and Greater Wellington These observed
differences between the three measures reflect differing
conceptualisations of ethnic segregation. In view of its better fit to
our outcome variable and the strong emphasis in previous work on the
suitability of the isolation index for studying segregation and
health, we chose to focus on the isolation index. In order to identify
possible non-linearities in the segregation-smoking relationship we
categorised the isolation index into quartiles.
Figure 3: Segregation Measures Compared
A B

C

We now consider whether the ecological UA-level relationship between
smoking and segregation is confounded by individual-level demography
or higher-level deprivation. Table 3 presents the results from models
A-D of the multilevel analysis, based on a reference category of a
woman aged 16-25 living in CAUs of average deprivation and an UA with
segregation in the lowest quartile. The odds ratios and associated
confidence intervals for the age and sex terms remain relatively
constant across the four models and the patterning between 1981 and
1996 is also consistent. Relative to the reference category, smoking
was raised in the 26-35 and 36-45 age groups in 1981, although this
was not statistically significant in the 26-35 group. In 1996 there
was a significantly raised probability of smoking only in the 26-35
group, In both years, the odds of smoking otherwise declined with age,
with the rate of decline being marginally greater in 1996. The odds of
a man being a smoker declined between 1981 and 1996. These results
give further confirmation of the smoking transition during the 1980s
and early 1990s and point clearly to prevalence concentrating among
younger people and cessation being a function of age and,
increasingly, sex.
Having considered demographic variation, we now turn to the effects of
deprivation and segregation on smoking bringing in the mean-centred
ecological deprivation measures and the quartile categorisation of the
isolation index. A comparison of the results from Model B with those
for Model D reveals differences in the odds ratios and confidence
intervals for the isolation indicators. In contrast, the deprivation
results between Models C and D vary little. This, alongside the
consistent results for the demographic terms in the two models
suggests that it is Model D rather than Model B, or the earlier
ecological analysis, that indicates the independent effect of
segregation on smoking. In the New Zealand context, deprivation is
confounding the effect of segregation on smoking.
The results of Model D suggest small but significant independent
effects on smoking in 1981 in UAs in the second and fourth quartiles
on the isolation index. This suggests that the impact of segregation
on smoking in 1981 was greater in UAs where the Māori population was
experiencing slightly more than minimal isolation or where it was most
isolated. Comparison with model B suggests that controlling for
deprivation attenuated previous indications of the isolation effect
across all levels of isolation and, in UAs with isolation in the third
quartile, the effect ceased to be significant. By 1996, isolation
effects had reduced further. Isolation in the highest quartile showed
hints of an association with lower smoking but, overall, no
statistically significant effects were identifiable: Māori isolation
at the UA level had ceased to have a discernible effect on smoking
over and above the impact of deprivation.
Table 3: Main Effects Multilevel Models: Odds Ratios and 95%
Confidence Intervals
Model A
Model B
Model C
Model D
Year
1981
1996
1981
1996
1981
1996
1981
1996
Base
% Smoking
33.314
28.470
30.683
26.173
32.805
31.110
30.938
30.798
Age 26-35
1.001
1.041
1.001
1.042
1.002
1.044
1.002
1.044
0.991 - 1.011
1.031 - 1.052
0.991 - 1.011
1.032 - 1.052
0.992 - 1.012
1.033 - 1.054
0.992 - 1.012
1.033 - 1.054
Age 36-45
1.011
0.901
1.011
0.901
1.011
0.900
1.011
0.900
1.001 - 1.021
0.892 - 0.910
1.001 - 1.021
0.892 0.910
1.001 - 1.021
0.891 - 0.908
1.001 - 1.021
0.891 - 0.908
Age 46-55
0.979
0.792
0.979
0.792
0.978
0.788
0.978
0.788
0.970 - 0.989
0.783 - 0.802
0.970 - 0.989
0.783 - 0.802
0.969 0.988
0.779 - 0.797
0.969 - 0.988
0.779 - 0.797
Age 56-65
0.780
0.563
0.780
0.563
0.777
0.555
0.778
0.555
0.771 - 0.790
0.556 - 0.571
0.771 - 0.790
0.556 - 0.571
0.768 - 0.786
0.548 - 0.563
0.769 - 0.787
0.548 - 0.563
Age 66-75
0.501
0.366
0.501
0.365
0.497
0.358
0.497
0.358
0.494 - 0.508
0.360 - 0.371
0.494 - 0.508
0.360 - 0.371
0.490 - 0.504
0.352 - 0.363
0.490 - 0.504
0.352 - 0.363
Age 75+
0.245
0.188
0.245
0.188
0.242
0.183
0.242
0.183
0.239 - 0.251
0.184 - 0.193
0.239 - 0.251
0.184 - 0.192
0.236 - 0.248
0.178 - 0.187
0.237 - 0.248
0.178 - 0.187
Male
1.220
1.100
1.220
1.100
1.223
1.103
1.223
1.103
1.213 - 1.227
1.093 - 1.106
1.213 - 1.227
1.094 - 1.107
1.215 - 1.230
1.096 - 1.109
1.215 - 1.230
1.096 - 1.109
Maori Isolation
Q2
Mid Low
1.220
1.042
1.146
1.029
1.100 - 1.354
0.892 - 1.216
1.051 - 1.249
0.954 - 1.102
Q3
Mid High
1.132
1.014
1.091
1.039
1.018 - 1.258
0.882 - 1.166
0.999 - 1.191
0.972 - 1.110
Q4
High
1.241
1.269
1.120
0.096
1.116 - 1.380
1.076 - 1.496
1.025 - 1.223
0.928 - 1.069
Deprivation
No Qualifications
1.011
1.018
1.012
1.019
1.009 - 1.013
1.016 - 1.020
1.010 - 1.014
1.017 - 1.021
No DPB
1.044
1.048
1.043
1.048
1.034 - 1.054
1.040 - 1.056
1.033 - 1.053
1.040 - 1.056
Unemployed
1.033
1.016
1.033
1.016
1.024 - 1.041
1.008 - 1.024
1.024 - 1.041
1.008 - 1.024
Renting
1.004
1.005
1.004
1.005
1.002 - 1.006
1.003 - 1.007
1.002 - 1.006
1.003 - 1.007
Model D also allows us to draw conclusions about the impact of
CAU-level deprivation on smoking in New Zealand. This impact generally
rose between 1981 and 1996, perhaps reflecting an increasing
concentration of smoking into deprived areas. The strongest
relationship was evident in CAUs with higher uptakes of domestic
purposes benefit. An exception to the strengthening impact of
deprivation on smoking was the association of smoking and UA
unemployment rates. This reduced and a tentative explanation would
reference the improving New Zealand economy over the period.
Significantly more complex models (not shown) were developed to
explore the possibility that isolation impacted differentially on
particular age-sex groups or in particular places. The results of
these additional analyses did not depart significantly from those
reported above. In 1981 isolation was on average associated with
smoking prevalences that were increased by two to three percentage
points across the (limited) range of observed levels of isolation. By
1996, this impact had fallen to less than one percent and was actually
associated with a (insignificant) reduction in smoking prevalence in
UAs with the highest levels of segregation.
These findings confirm previous work on area effects on smoking both
in New Zealand and elsewhere (Barnett et al., 2004; Duncan et al.,
1999) which has found that contextual effects are small. Neither
deprivation nor isolation are as important in understanding social and
spatial variations in smoking prevalence as age and sex. Overall we
conclude that the UA impact of isolation on smoking is very small and
it is as likely to attenuate estimated smoking prevalence as it is to
increase it. Isolation effects are also significantly confounded in
the New Zealand context, by deprivation. With the size of the effects
that are involved we do not believe further speculation about pathways
or processes is appropriate, but it would be tempting to hypothesise
that communities with the highest levels of isolation may, in certain
circumstances, possess internal norms that, to an extent, protect them
from the full impact of deprivation on smoking prevalence.
Conclusions
===========
Null results are an important but often hidden aspect of scientific
inquiry, potentially contributing as much to knowledge as
superficially more successful studies that prove hypotheses and
provide positive advances to understanding. The under-reporting of
null results has long been known (Sackett, 1979). The research
presented in this paper has undoubtedly uncovered a (largely) null
result but it has also moved the study of the impact of ethnic spatial
segregation on health forward in two ways: first by moving the focus
of study from health outcomes to health-related behaviour and, second
by working in a setting where segregation is less stark than the
prevailing US location of most previous work. We have additionally
contributed to an ongoing debate about the extent to which
relationships between segregation and health are confounded by
deprivation and, following methodological strictures in previous work,
we have explored different measures of segregation and ensured that
our study has been conducted in an appropriate multilevel framework.
In terms of a theoretical contribution, we have considered how and to
what extent segregation plays a part in psychosocial explanations of
health-related behaviour.
Some limitations to our study must be acknowledged. Though the New
Zealand census provides a significant resource for research on
smoking, there is an estimated non-response to the smoking questions
of c.7% (Statistics New Zealand, 2007). Undoubtedly this
non-response is non-random with respect to age, sex, deprivation and
geography. Second, our data are of course repeated cross sections. Our
census data do not allow us to trace the impact of segregation on
particular individuals through time. Third, the variation in our
exposure measures (residential segregation), though clear, is not
great. This may have impacted on the robustness of our conclusions
though it is clearly reflective of the New Zealand situation. Fourth,
as we noted earlier, confidentiality constraints meant that we did not
have access to four-way census cross-tabulations allowing us to
consider ethnic smoking outcomes, controlled for age and sex. We
confronted this limitation by limiting our focus to the impact of
segregation on age-sex smoking rates. The pathways for this
generalised impact are clear (segregation would be expected to raise
smoking by both minority and majority populations though psychosocial
stress) and they have sound foundations in past work. We acknowledge
that, in a variant of the ecological fallacy, the limited generalised
impact of segregation on smoking may mask specific impacts on smoking
among different ethnic groups.
We hope to develop this research further. It is our intention to
consider the potential relationship of ethnic segregation to other
measures of smoking behaviour, notably measures of reduction and
cessation. We are also continuing to investigate alternative
approaches to modelling ethnic-specific smoking behaviour in relation
to segregation and propose to test more localised intra-urban measures
of segregation rather than following the current view of segregation
as a construct most appropriately considered at the level of the urban
area as a whole. This may also enable us to test associations in
models where residential segregation is less invariant. Research
attaching area segregation measures to recent survey data is a further
possibility, and we anticipate developing different study designs,
particularly in relation to the temporal evolution of the spatial
relationship between segregation and smoking. Finally we intend,
through a combination of innovation in study design and careful
negotiation of challenges to data confidentiality, to model
age-sex-ethnicity smoking prevalences in order to consider the
differential effect of isolation on, for example, Māori smokers.
Our overall conclusions are fourfold. First, we contend that our
chosen measure of segregation, Māori isolation, has only a limited
relevance to smoking behaviour in New Zealand. This challenges some
previous work on health topics conducted in the USA and perhaps
reflects the lower levels of segregation in New Zealand. It also
eventuates from the ecological rather than multilevel design of some
past studies in which the composition of areas was not taken into
account. Second, it is generally the case that deprivation confounds
initial indications of a relationship between smoking and ethnic
spatial segregation. In terms of our conceptual model, segregation
influences smoking via deprivation but only weakly. The increasing
isolation experienced between 1981 and 1996 in New Zealand UAs was
overshadowed by deprivation as a contextual determinant of smoking
behaviour. Third, where weak segregation effects could be discerned,
they exhibited complex patterns, sometimes protecting from the effect
of deprivation, sometimes enhancing that effect, sometimes operating
only at higher levels of segregation, sometimes also at relatively low
levels. Finally, and in relation to the implications of our findings
for policy, it is important to end on a note of caution. This study
suggests that Māori isolation is not a major factor in identifying
high smoking communities. It must be emphasised that this conclusion
is about a specific measure: Māori isolation. It is not about the
clear and incontrovertible link between Māori and higher levels of
smoking. It does not lessen the need for Māori-specific smoking
policies and services while stark inequalities in smoking behaviour
between Māori and Europeans in New Zealand persist.
Acknowledgements
We thank Statistics New Zealand for providing the census data used in
this paper..
References
==========
1 The term ‘smoking transition’ refers to the temporal shift of higher
smoking prevalences from higher to lower status groups, men to women
and more developed to less developed countries.
2 The New Zealand urban area definitions identify separate areas
within Greater Auckland and Greater Wellington. These areas correspond
to natural divisions and we have retained them in our analysis.
3 The formulae for these indices are widely available and may be
obtained from the corresponding author.