smoking and deprivation: are there neighbourhood effects?
TRANSCRIPT
Smoking and deprivation: are there neighbourhood e�ects?
Craig Duncan *, Kelvyn Jones, Graham Moon
Institute for the Geography of Health, Department of Geography, University of Portsmouth, Buckingham Building, Lion Terrace,
Portsmouth PO1 3HE, UK
Abstract
Debate has centred on whether the character of places plays an independent role in shaping individual smokingbehaviour. At the small-area scale, particular attention has focused on whether measures of neighbourhood
deprivation predict an individual's smoking status independent of their own personal characteristics. This studyapplies multilevel modelling techniques to data from the British Health and Lifestyle Survey and ward (localneighbourhood) level deprivation scores based on four variables from the national Census. Results suggest that after
taking account of a large range of individual characteristics, both as main e�ects and interactions, together withcomplex structures of between-individual variation, measures of neighbourhood deprivation continue to have anindependent e�ect on individual smoking status. In addition, signi®cant between-ward di�erences in smoking
behaviour remain which cannot be explained either by population composition or ward-level deprivation. The studysuggests that the character of the local neighbourhood plays a role in shaping smoking behaviour. # 1998 ElsevierScience Ltd. All rights reserved.
Keywords: Smoking; Deprivation; Neighbourhood e�ects; Multilevel modelling
1. Introduction
Over the last two decades a series of empirical stu-
dies have identi®ed signi®cant geographical variations
in health-related behavioural practices in Britain
(Cummins et al., 1981; Balarajan and Yuen, 1986;
Dunbar and Morgan, 1987; Braddon et al., 1988;
Richards, 1989; Blaxter, 1990; Whichelow et al., 1991).
In most of this work, attention has focused on di�er-
ences at the regional level with only a few studies
examining more local in¯uences (Cummins et al., 1981;
Richards, 1989; Blaxter, 1990). Similar research has
been conducted in America with signi®cant area-based
di�erences in health-related behavioral practices being
identi®ed (Marks et al., 1985; MMWR, 1987; Hilton,
1988; Klein and Pittman, 1993; Colby et al., 1994). As
is now well-recognised, geographical variations in out-
come measures do not necessarily mean that places
have signi®cance in their own right. Variations may
arise due to particular types of people, whose personal
characteristics are closely linked to the outcome under
study, being found more commonly in certain places.
Thus, geographical variations may not re¯ect indepen-
dent area e�ects; they may simply be the result of
spatially di�erentiated population composition.
Given the aetiological signi®cance now attached to it
(Doll and Peto, 1981; Royal College of Physicians,
1983; Department of Health, 1992), it is not surprising
that nearly all studies of geographical variations in
health-related behaviour in Britain have considered
smoking. The majority of this work has suggested,
either implicitly or explicitly, that variations in
Social Science & Medicine 48 (1999) 497±505
0277-9536/98/$ - see front matter # 1998 Elsevier Science Ltd. All rights reserved.
PII: S0277-9536(98 )00360-8
PERGAMON
* Corresponding author. Tel.: +44-1705-842-507; fax: +44-
1705-842-512; e-mail: [email protected].
smoking behaviour across Britain are not simply an
artefact of varying population compositions but are,additionally, the result of independent `area' or `con-textual' e�ects. This position is well exempli®ed by M.
Blaxter's major piece of analytical, academic research,Health and Lifestyles, in which she concludes that``class is related to smoking in di�erent ways in di�er-
ent types of areas'' (Blaxter, 1990, p. 117). Blaxter'sanalysis does, however, only involve the calculation of
area-based rates standardised for a limited number ofpersonal characteristics. Whilst other studies haveadopted a more sophisticated regression modelling
strategy, they have tended to rely on traditional single-level techniques. Such techniques can be a�ected by
serious problems of mis-estimation associated withauto-correlation since people living in the same placecan be expected to be more alike than a random
sample (Scott and Holt, 1982; Aitken and Longford,1986; Skinner et al., 1989). In addition, they are alsoprone to technical problems associated with small
numbers and sampling ¯uctuations (Jones, 1991).Multilevel regression modelling techniques
(Goldstein, 1995) have been shown to o�er a moresuitable way of conducting analyses of area e�ects(Jones and Duncan, 1996). In the United States, such
an approach has found signi®cant area e�ects forsmoking behaviour with people in the highest preva-
lence communities being ``about 7 percentage pointsmore likely to smoke'' after taking account of person-level variables (Diehr et al., 1993, p. 1147). In Britain,
initial work based on similar techniques has focusedmore on the regional level rather than smaller geo-graphical scales (Duncan et al., 1993). Results
suggested that, in comparison with individual-levelcompositional e�ects, neither wards nor regions play a
signi®cant independent role. The models used in thisanalysis were, however, rather limited. More recentwork in Scotland by Hart et al. (1997) has obtained
similar results at the local scale. In this study, smokingwas found to be the coronary heart disease risk factormost closely related to individual-level compositional
e�ects. Indeed, after taking account of these, thedegree of contextual variation in smoking behaviour
was minimal. While this analysis did usefully considergender di�erences in detail (though, signi®cantly, nonewere found for smoking) the models used were, once
again, relatively simple. Furthermore, the study wasbased on only 22 local government districts.
In a slightly earlier piece of work, more complexmultilevel models have been applied in a study focus-ing on the North West Thames Regional Health
Authority in England (Kleinschmidt et al., 1995).Interestingly, this work did not originate from thedebate concerning the relevance of `place' for smoking
behaviour but from a desire to control for the possibleconfounding e�ect that smoking may have in ecologi-
cal studies examining the health e�ects of point sourcesof environmental pollution. Since smoking data are
usually unavailable in such studies, researchers havebeen interested in examining whether small area depri-vation indicators o�er a suitable surrogate for predict-
ing smoking behaviour. Through models including ahigher-level variable (Jones and Duncan, 1995), thisstudy found that measures of deprivation at the elec-
toral ward level were valid predictors of individualsmoking behaviour. Most importantly, results showedthat the level of deprivation of the immediate area of
residence remained a signi®cant predictor of smokingstatus after taking into account the age, gender andsocio-economic group of individuals.The present study seeks to contribute further to
work in this area by using a multilevel approach toexamine ward-level e�ects for smoking and deprivationusing data from a nationally representative survey
rather than just one regional health authority. LikeKleinschmidt et al. (1995), such e�ects are measuredthrough the inclusion of a higher-level composite vari-
able based on Census data. In contrast, however, thepresent study considers a broader range of individualsocio-structural characteristics (e.g.: housing tenure,
employment status, educational status as well as socialclass) rather than just a single detailed socio-economicgroup classi®cation. In addition, a wide range of indi-vidual-level interactions are also explored. This is im-
portant since, unless this is done, it is always possibleto argue, like Hauser (1970), that apparent contextuale�ects are simply the result of the mis-speci®cation of
individual e�ects. The present study also extends recentwork by applying other, more complex models whichtake into account di�erences between di�erent types of
people in their degree of variability. Again, this hasbeen shown to have an important confounding in¯u-ence on estimates of area e�ects (Bullen et al., 1997).Finally, unlike some earlier work, the results reported
here are based on recently improved estimation pro-cedures.
2. Methods
The data used in this analysis were obtained fromthe Health and Lifestyle Survey (Cox et al., 1987). Thesurvey focuses on the ``lifestyles, behaviours and cir-
cumstances relating to the physical and mental healthof the population'' (p. 1) and is recognised as beingone of the most comprehensive studies of factors per-
taining to the health of the adult British population todate. Although not conceived as a longitudinal panelstudy, additional funding meant that respondents to
the original survey conducted in 1984/1985 (HALS1)could be reinterviewed in 1991/1992 (HALS2). As sub-stantially fewer people were included in the repeat
C. Duncan et al. / Social Science & Medicine 48 (1999) 497±505498
exercise (5352 compared to 9003) we have used data
from the original survey, HALS1, in order to ensure
that the present study comes as close as possible in
meeting the guidelines for sample sizes in multilevel
analysis of Paterson and Goldstein (1991)1. While
other more recent surveys include data on smoking
behaviour (e.g. The Health Survey for England, (White
et al., 1993) and the General Household Survey
(Smyth and Brown, 1992)), none at present, unlike
HALS1, allow the speci®c residential location of
respondents to be identi®ed making it impossible to
include measures of ward deprivation as is required
here. Two ®nal justi®cations for using HALS1 is that
such data formed the basis of Blaxter's important
work discussed earlier and results here will provide a
useful comparative baseline should other, more suit-
able large-scale surveys be undertaken in the future.
Since the survey was based on a three stage
sampling design Ð individuals (9003) within wards
(396) within constituencies (198) Ð it is ideally suited
to a multilevel approach. Previous research examining
ward-level e�ects on health outcomes using data from
HALS1 has been restricted to working with a subset of
wards as not all wards could be matched with those
used in the 1981 Census (Humphreys and Carr-Hill,
1991). As part of the present research, however, com-
plete matching was achieved by using the individual
postcodes of each respondent which were obtained
from the survey's principal investigator2.Having completed the identi®cation procedure, an
index of deprivation was derived for each ward using
data from the 1981 Census. Four variables were used:percentage of economically active population who are
seeking work; percentage of people in householdswhere the home is rented from the Local Authority;percentage of people in households where the head of
household is in social class IV/V; percentage of peoplein households without access to a car. Each of thesevariables was ranked and substituted with an equiva-
lently ranked normal score and the ®nal index wasformed from the sum of these ranked normal values.
Such a procedure ensures that the index is not undulya�ected by a highly non-normal distribution for anyparticular census variable. The ®nal index can be taken
to represent the broad socio-economic ecology of thearea in which respondents lived and scores rangedfrom ÿ9.2 for the least deprived area (Northaw at
Welwyn and Hat®eld in Hertfordshire) to 10.3 for themost deprived (Hulme in Manchester).
As stated earlier, it is essential that individual e�ectsare well-speci®ed if contextual e�ects are to be esti-mated accurately. In order to try and achieve this, an
ordinary single-level analysis was ®rst carried out usingthe S-Plus software package. As the response consistedof a binary variable classifying individuals as either
being regular smokers or not, models based on a logitlink function were used. Regular smokers were de®ned
as those who responded to a question concerning theirdaily cigarette consumption by declaring a rate of onecigarette or more. Variation in this response was re-
lated to a series of explanatory variables re¯ecting arange of individual characteristics. Besides age, genderand social class, as used by Kleinschmidt et al. (1995),
housing tenure, employment status, educational statusand marital status were also taken into account as pre-
vious research has shown that these are important pre-dictors of smoking status (Hunt et al., 1988; Pierce,1989; Pugh et al., 1991; Oakley, 1994). As well as
including the main e�ects, all two-way interactionswere explored between the following variables: gender,social class, tenure and education.
The fundamentals of multilevel modelling as itapplies to health research have been covered compre-
hensively elsewhere (Rice and Leyland, 1996). The pre-sent study focused on a three level analysis of theHALS1 data with individuals at level 1 nested within
electoral wards at level 2 and regions, based on TheEconomist classi®cation, (Johnston et al., 1988) at
level 3. Regions were included to re¯ect the possibilityof higher-level contextual e�ects although it is at theward level that results are explored in detail. For the
reason outlined above, logistic multilevel models basedon a logit link function were used (Goldstein, 1991).
1 In HALS1, the average number of individuals per ward is
22.375 (S.D. = 3.35). Obviously, this is much higher than in
HALS2 and is much closer to meeting Paterson and
Goldstein's recommended suggestion of 25. Since we are
using such a large number of places (see later), it is believed
that the present analysis does provide realistic and valuable
results.2 These postcodes were associated with grid references
through geographical information systems technology avail-
able at the University of Manchester Computing Centre.
Research has shown that this procedure, which uses the
Central Postcode Directory, can be prone to error and an
improvement process has been suggested involving the simple
addition of 50 ms to the x and y coordinates returned
(Gatrell et al., 1991). This procedure was followed. The `point
in polygon' technique available within GIS technology was
then used to associate the postcodes with a set of digitised
ward boundaries from the 1981 Census (ESRI, 1993). Due to
the spatial coarseness of the grid-referencing, generalisation in
the ward boundary ®les, and the lack of coincidence between
the 1981 and the 1984/1985 boundaries, it was found that
mis-allocation could occur. Consequently, the results were
examined in detail and respondents were allocated to the 1981
ward in which the majority of them with the same HALS1
identi®er were found to be located. An indicator variable was
included in the models to identify wards that proved di�cult
to identify (22 of the 396). Analyses were undertaken with
and without this indicator variable and no substantial nor sig-
ni®cant di�erences were found.
C. Duncan et al. / Social Science & Medicine 48 (1999) 497±505 499
Models were ®tted which included the individual-level
main e�ects on their own and then together with any
signi®cant two-way interactions found in the single-
level analysis. Apart from age, which was represented
as a continuous variable centred about its mean, the
main e�ects were represented by a set of dummy, indi-
cator variables which were contrasted with the base
category of a 46-year old employed woman who left
school before the age of 16, is married and lives in an
owner occupied household, the head of which is in
social class III-manual. This base category represents
the set of individual characteristics that occur most fre-
quently amongst the 9003 respondents to the question-
naire Ð that is, the stereotypical individual. In
addition to including the individual-level e�ects, the
deprivation scores for each ward were also included as
a higher-level continuous variable (Jones and Duncan,
1995)3.
Any variation remaining after taking account of the
individual and ward level characteristics is apportioned
to the appropriate level through the inclusion of ran-
dom terms at each level of the model. At the individual
level, extra-binomial variation was allowed for
(Collett, 1991). Research has shown that in multilevel
logistic models the results for the higher level random
e�ects can be seriously underestimated (Breslow and
Clayton, 1993; Rodriguez and Goldman, 1995). In the
present analysis, the software package MLn was uti-
lised which allows users to select estimation procedures
based on the second-order Taylor expansion and par-
tial quasi-likelihood approximations (Rasbash and
Woodhouse, 1995). As these have been shown to be
least a�ected by problems of underestimating higher-
level random terms they were used in all of the models
reported here (Goldstein, 1994). Importantly, since
such procedures were not available at the time of the
original multilevel analysis (Duncan et al., 1993) the
present analysis will con®rm whether the conclusionsof this earlier British work are reasonable4.
3. Results
3.1. Individual interaction e�ects
As outlined, a single-level analysis was ®rst per-
formed to establish whether there were any signi®cantinteraction e�ects. Of all the interaction terms ®tted,only those between male and missing tenure and localauthority renter and `other' social class were found to
be signi®cant. Multilevel analyses were conducted bothwith and without these terms. Signi®cantly, no import-ant di�erences were found in terms of the substantive
interpretation of neighbourhood e�ects. This ®ndingcasts doubt on suggestions that neighbourhood e�ectsare simply the result of mis-speci®ed individual level
models neglecting interaction e�ects. Consequently, theresults presented in the following subsections are forthe simpler main e�ects only multilevel models5.
3.2. Individual main e�ects
Table 1 shows results for models in which predictor
variables were included and any remaining variationwas apportioned to a single intercept term at eachlevel. The ®rst four columns are for a model in which
only the individual-level predictors are included; thelatter four for when the ward-level deprivation par-ameter is also included. In both cases, the estimates, as
well as their standard errors and corresponding Zscores are given. In addition, odds ratios derived asthe exponential of the estimated coe�cients are givenfor the categorical predictors.
If we consider the model with only individual-levelpredictors, we can see a number of important featuresconcerning the patterning of smoking behaviour on the
basis of individual characteristics as indicated by thelevel-1 ®xed e�ects. To interpret these results it mustbe remembered that they represent contrasts on each
variable from the base categories that characterise thestereotypical individual. For an estimate to be signi®-cantly di�erent from zero at the 0.05 level, the Z rationeeds to be more than22. In accordance with previous
research, men are more likely to be smokers thanwomen and the probability of smoking decreases asage increases. Besides age and gender di�erences, we
see strong patterning on the basis of the other individ-ual characteristics with the chances of an individualbeing a smoker being signi®cantly higher should they
be in the lower social class categories, live in rented ac-commodation, have fewer years of schooling, be out ofwork or divorced or separated.
3 It is important to note that including place characteristics
as higher-level variables in multilevel models provides a more
accurate estimation of their explanatory power than if they
are included within traditional single-level models (Aitken and
Longford, 1986). A full technical account of this, together
with several empirical illustrations, can be found in Bryk and
Raudenbush (1992).4 It should be noted that these improved estimation pro-
cedures were applied in the later work of Hart et al. (1997) in
Scotland (pers. comm).5 It should be noted that analyses were also undertaken
which included ethnicity and household income as main
e�ects whilst nonlinear relationships with age and age±gender,
age±tenure, age±education and gender±unemployment inter-
actions were also considered. Once again, the ®ndings with
regards to neighbourhood e�ects were neither substantially
nor signi®cantly di�erent.
C. Duncan et al. / Social Science & Medicine 48 (1999) 497±505500
3.3. Neighbourhood (ward) e�ects
Turning our attention to the model that also
includes the level-2 ®xed e�ect, deprivation, we can see
that the parameter estimate is statistically signi®cant.
Furthermore, including this variable leads to a sub-
stantial reduction in the likelihood (w 2=22, p < 0.01).
Given the inclusion of the other variables in the
model, this shows that ward-level deprivation does
seem to have an independent e�ect on individual
smoking status. Interestingly, the size of the e�ect for
the index used here is virtually identical to that
Table 1
Results for main e�ects models excluding and including ward-level deprivation
Excluding ward-level deprivation Including ward-level deprivation
estimate standard Z odds estimate standard Z odds
(logits) error ratio ratio (logits) error ratio ratio
Fixed e�ects
Level 1 (individual)
Intercept ÿ0.74 0.06 ÿ12.33 ÿ0.74 0.06 ÿ12.33Age ÿ0.02 0.002 ÿ10.00 ÿ0.02 0.001 ÿ20.00
Gender
Male 0.14 0.05 2.80 1.15 0.14 0.05 2.80 1.15
Social class
I&II ÿ0.32 0.07 ÿ4.57 0.73 ÿ0.29 0.07 ÿ4.14 0.75
III nonmanual ÿ0.30 0.08 ÿ3.75 0.74 ÿ0.28 0.08 ÿ3.50 0.76
IV&V 0.04 0.06 0.67 1.04 0.04 0.06 0.67 1.04
Missing/other ÿ0.32 0.18 ÿ1.78 0.73 ÿ0.29 0.18 ÿ1.61 0.75
Employment status
Unemployed 0.51 0.11 4.64 1.67 0.49 0.11 4.45 1.63
Housing status
Local authority renter 0.68 0.06 11.33 1.97 0.61 0.06 10.17 1.84
Other renter 0.49 0.09 5.44 1.63 0.47 0.09 5.22 1.60
Missing 0.89 0.40 2.23 2.44 0.85 0.41 2.07 2.34
Marital status
Single ÿ0.19 0.07 ÿ2.71 0.83 ÿ0.19 0.07 ÿ2.71 0.83
Widowed ÿ0.18 0.10 ÿ1.80 0.84 ÿ0.19 0.10 ÿ1.90 0.83
Divorced/separated 0.49 0.10 4.90 1.63 0.49 0.10 4.90 1.63
Age leaving school
16 ÿ0.41 0.07 ÿ5.86 0.66 ÿ0.40 0.07 ÿ5.71 0.67
Post-16 ÿ0.58 0.07 ÿ8.29 0.56 ÿ0.56 0.08 ÿ7.00 0.57
Missing ÿ1.59 0.63 ÿ2.52 0.20 ÿ1.59 0.63 ÿ2.52 0.20
Fixed e�ects
Level 2 (ward)
Deprivation 0.04 0.01 4.00
Random e�ects variance
Level 3 (region)
Intercept 0.011 0.008 0.002 0.004
Level 2 (ward)
Intercept 0.051 0.02 0.045 0.02
Level 1 (individual)
Intercept 0.992 0.02 0.993 0.02
C. Duncan et al. / Social Science & Medicine 48 (1999) 497±505 501
obtained in the North West Thames study for theCarstairs index (Carstairs and Morris, 1989). In terms
of the present analysis, the odds ratio of the stereotypi-cal individual described earlier being a smoker for the95% percentiles on the area deprivation index
(ÿ6.5401; 6.1946) is 1.66 (95% C.I. 1.29, 2.15).Inspecting the odds ratios given in Table 1 shows thatthis is comparable to most of the e�ects for important
individual level explanatory variables.Further work was conducted to investigate the
nature of this e�ect in more detail. First, a series of
random coe�cient models were estimated to assesswhether the coe�cients for any of the individualcharacteristics varied between wards. Such modelsallow for the possibility that people with di�erent indi-
vidual characteristics may be di�erentially likely to besmokers depending on where they live. The resultsshowed no signi®cant di�erences for any of the charac-
teristics. In terms of social class characteristics thiscon®rms the ®ndings of the North West Thames studyand is in contrast to Blaxter's conclusions: people of
di�erent social classes did not appear to display di�er-ent degrees of between-neighbourhood variability inthe probability of being a smoker. Models were also
estimated to assess whether the e�ect of neighbour-hood deprivation was di�erent for di�erent types ofpeople. This was done through the inclusion of cross-level interaction variables as described in detail by
Jones and Duncan (1995). While there was some sug-gestion that the e�ect of neighbourhood deprivationwas reduced for local authority renters and those who
left school at 16, no other di�erences were found.Given these results, another model was estimated toexamine if there was any suggestion that a general
deprivation e�ect could be nonlinear. No evidence wasfound to support this. Lastly, a model was estimatedin which a general deprivation coe�cient was allowedto be random at the region level. Again, no signi®cant
variation was found, suggesting that the e�ect ofneighbourhood deprivation on individual smoking sta-tus is the same across all regions of the country. By
way of summary, it would appear from all of theseresults that neighbourhood deprivation does have agenuinely independent e�ect upon smoking behaviour,
but it is an e�ect which is not, for the large part,socially or geographically variable.
3.4. Between-individual variation
As multilevel research in other areas has shown,
complex forms of between-individual variation can beimportant, both in their own right and as a confound-ing factor (Goldstein, 1995; Bullen et al., 1997). As sta-
ted above, extra-binomial variation was allowed for atthe individual level. All of the previous models have,however, only allowed for under- and overdispersion
in the sample as a whole, yet it is possible that suche�ects are speci®c to particular types of people. Failing
to take such complex structures of between-individualvariation into account has been shown to be respon-sible for arti®cially exaggerating the size of area e�ects
in the case of continuous response data (Bullen et al.,1997). To consider whether this may also apply here inthe case of discrete response data, a series of ®nal
models were estimated in which indicator variables forage, gender and social class were included in the level-1 random part. No signi®cant degree of over- or
underdispersion was found in any of these models.This suggests that the results in Table 1, column (b)can be thought of as representing the ®nal model.
3.5. Between-neighbourhood variation
Contrary to previous work (Diehr et al., 1993;
Duncan et al., 1993), it is not possible to decomposethe remaining variation into level 1 and level 2 pro-portions as the estimates are based on di�erent
measurement scales (Rasbash, pers comm). Table 1shows, however, that after including both individualand ward predictors the ward-level random term
remains fairly large and is, in fact, still signi®cant(w 2=5.40, p = 0.02). Thus, neither the level of neigh-bourhood deprivation nor di�erences in populationcomposition explain all of the between-ward variability
in individual smoking status. Obtaining the posteriorestimates for each particular neighbourhood showsthat this `unexplained' variation amounts to a gap of 8
percentage points in the probability of individualsbeing regular smokers between the 95% percentiles forthese random e�ects. Whilst this is reasonably large,
the con®dence intervals associated with each of theseestimates show that there is a high degree of overlapdue to the large degree of uncertainty associated withthem. Thus, while in the population at large there are
signi®cant di�erences remaining between wards, identi-fying particular neighbourhoods as being either highor low smoking places is problematic and cannot prop-
erly be attempted.
4. Discussion and conclusions
The results presented here, based on data from a
national survey, support the recent suggestion thatthere are area e�ects on smoking behaviour whichrelate to the level of deprivation of the immediate area
of residence of individuals. Importantly, the presentanalysis recognises a broader range of personal charac-teristics than previous work both as main e�ects and
as interactions. At the same time, complex structuresof between-individual variation have also been takeninto account. Hence, the present work can be thought
C. Duncan et al. / Social Science & Medicine 48 (1999) 497±505502
of as giving added weight to the suggestion that neigh-
bourhood deprivation has a genuinely independente�ect and that it is not simply an artefact of popu-lation composition.
In substantive terms, these ®ndings con®rm the con-tention of Kleinschmidt et al. (1995), that small area
deprivation measures may be used to correct for thepossible confounding e�ect of smoking in ecologicalstudies examining the relationship between point
sources of environmental pollution and health. AsKleinschmidt et al. point out, however, census-basedsmall area deprivation measures are rather crude and
do not summarise variations in smoking prevalence aswell as corresponding individual measures. Thus, while
using deprivation indicators as surrogate variables inpollution studies does make sense, they are not a par-ticularly well-re®ned measure. To improve on this, one
alternative would be to combine estimates of both areaand individual e�ects derived from modelling nationalsurvey data with appropriate small area data from the
population Census. Procedures for producing such syn-thetic estimates of smoking prevalence rates is cur-
rently being developed in other work. Full details canbe found in Twigg et al. (1998).The ®ndings here also have relevance for the wider
debate concerning the importance of independent areae�ects on health-related behaviour, or, more speci®-cally, smoking status. This study has demonstrated
that places, at the neighbourhood level, are related tosmoking behaviour independent of the type of people
that they contain. Although it is not possible in astudy such as this to be precise about the mechanismsand processes by which neighbourhood characteristics
in¯uence individual smoking behaviour, the resultsobtained do provide some useful pointers. First, it isapparent that the contextual e�ects operate in the
same `direction' as the individual e�ects. That is, lowindividual status is characterised by a greater chance
of being a smoker as is living in a low status place. Inthe political science literature, this is known as a `con-sensual environmental e�ect' (Jones and Duncan,
1995, p. 32). Such a description connects with therecent typology of spatial e�ects in work on healthoutcomes of Macintyre (1996). As she points out, one
way in which place can matter is through a `socialmiasma' e�ect; that is, `collective' group properties
exert some in¯uence over and above individual proper-ties. Hence, in the present case, as you go into lowerstatus areas the behavioural characteristics of lower
status people are reinforced and dominate those ofothers.
What is also apparent from the present research,however, is that neighbourhood characteristics tend tohave a general rather than a socially speci®c e�ect.
With the exception of local authority renters and thosewho left school at 16, the e�ect of area deprivation
upon smoking behaviour is uniform across di�erent
types of people. It may be the case, therefore, that ad-ditionally (or, perhaps, alternatively) place may matterthrough what Macintyre describes as a `contextual'
e�ect. Thus, living in deprived areas has an impact onbehaviour not (simply) because of the presence ofmore smokers but (also) due to living in unpleasant,
undesirable, unsafe environments in which there arefewer opportunities for making `healthy choices'
(Sooman and Macintyre, 1995; Macintyre, 1996).Obviously, untangling these two possibilities is a
complex task. Indeed, it might be argued that the latter
should be thought of as producing, at least in part, theformer. Notwithstanding this, it may be possible that
in similar future work the inclusion of a range ofhigher-level predictors, some re¯ecting social collectiv-ities, others re¯ecting environmental quality, may go
some way to disentangling the di�erent types of e�ects.For the moment, what is clear is that the social en-vironment in which people live and lead their lives is
of signi®cance. As argued elsewhere, therefore, improv-ing health should not simply involve targeting individ-
uals (McKinlay, 1994; Syme, 1996; Marmot, 1998).Besides these substantive ®ndings, we would also
like to emphasise one important methodological point.
In studies such as this, it is always possible to arguethat independent place e�ects are simply an artefactcaused by the mis-speci®cation of individual-level
e�ects. In the present case, we would contend thatsuch an argument is invalid given the wide range of in-
dividual characteristics taken into account togetherwith the fact that interaction e�ects were also con-sidered. More importantly, we would further contend
that such an argument needs to be routinely appliedthe other way round. By this, we mean to suggest thatin the same way that place e�ects may be exaggerated
if important individual characteristics are not included,the signi®cance of individual e�ects may be overesti-
mated if place characteristics are excluded. This can beexpected, at least in the British context, since asMacintyre et al. (1993) emphasise, and as we have
already suggested, certain types of people are morelikely to live in certain types of area. Thus, if import-ant higher-level variables are excluded which are corre-
lated with individual level variables that are included,models can be expected to overstate the signi®cance of
individual e�ects.Some evidence in support of this can be seen in the
present case by comparing the results in Table 1 for
housing tenure and social class before and after includ-ing the ward-level predictor. What is most important
to realise, however, and which actually does not applyhere, is that if higher-level variance terms are estimatedas being nonsigni®cant after including individual-level
variables, it does not necessarily follow that higher-level variables should be ignored. Omitted variable
C. Duncan et al. / Social Science & Medicine 48 (1999) 497±505 503
problems may, therefore, apply just as much at thecontextual level as at the individual one. Given this, it
is possible that had higher-level predictors beenincluded in the earlier work of Hart et al. (1997) sig-ni®cant (®xed) place e�ects for smoking may have
been found.The present study has also shown that (random)
place e�ects exist which cannot be `explained' either by
their population composition or their level of depri-vation. As discussed, the type of analysis conductedhere does, in theory, provide a means of identifying
particular neighbourhoods which exhibit either high orlow rates of smoking occurrence having taken accountof these factors. Since this would o�er guidance forcommunity health promotion e�orts, such an exercise
would undoubtedly be of interest to those working inpublic health. As with recent work on health servicesperformance (Goldstein and Spiegelhalter, 1996) it has,
however, been suggested that, in practice, very realproblems exist with such an identi®cation proceduredue to the degree of uncertainty associated with the
estimates.In conclusion, the present study has shown that the
level of deprivation in the immediate neighbourhood
of residence does have a signi®cant independent e�ecton individuals' smoking status. At the same time,other unidenti®ed factors associated with the characterof the local neighbourhood also play some role.
Acknowledgements
We would like to thank the ESRC Data Archive at
the University of Essex for providing the Health andLifestyle survey data (Cox, 1988), Professor Brian Coxfor supplying the individual postcodes, Steve
Frampton for much hard work on the geo-referencingand three anonymous referees for very helpful com-ments.
References
Aitken, M.A., Longford, N.T., 1986. Statistical modelling
issues in school e�ectiveness studies. Journal of the Royal
Statistical Society (Series A) 149, 1±26.
Balarajan, R., Yuen, P., 1986. British smoking and drinking
habits: regional variations. Community Medicine 8, 131±
137.
Blaxter, M., 1990. Health and Lifestyles. Tavistock/
Routledge, London.
Braddon, F.E.M., Wadsworth, M.E.J., Davies, J.M.C.,
Cripps, H.A., 1988. Social and regional di�erences in food
and alcohol consumption and their measurement in a
national birth cohort. Journal of Epidemiology and
Community Health 42, 341±349.
Breslow, N.E., Clayton, D.G., 1993. Approximate inference in
generalised linear mixed models. Journal of the American
Statistical Association 88, 9±25.
Bryk, A.S., Raudenbush, S.W., 1992. Hierarchical Linear
Models: Applications and Data Analysis Methods. Sage,
Newbury Park.
Bullen, N., Jones, K., Duncan, C., 1997. Modelling complex-
ity: analyzing between-individual and between-place vari-
ation: a multilevel tutorial. Environment and Planning A
29, 585±609.
Carstairs, V., Morris, R., 1989. Deprivation: explaining di�er-
ences in mortality between Scotland and England and
Wales. British Medical Journal 299, 886±889.
Colby, J.P., Linsky, A.S., Straus, M.A., 1994. Social stress
and state-to-state di�erences in smoking and smoking re-
lated mortality in the United States. Social Science and
Medicine 38, 373±381.
Collett, D., 1991. Modelling Binary Data. Chapman and
Hall, London.
Cox, B.D., 1988. Health and Lifestyle Survey, 1984±1985
(computer ®le). ESRC Data Archive, Colchester.
Cox, B.D., Blaxter, M., Buckle, A.L.J., et al., 1987. The
Health and Lifestyle Survey: a Preliminary Report. Health
Promotion Research Trust, London.
Cummins, R.O., Shaper, A.G., Walker, M., Wale, C.J., 1981.
Smoking and drinking by middle-aged British men: e�ects
of social class and town of residence. British Medical
Journal 283, 1497±1502.
Department of Health, 1992. The Health of the Nation: a
Strategy for Health in England. HMSO, London.
Diehr, P., Koepsell, T., Cheadle, A., Psaty, B.M., Wagner, E.,
Curry, S., 1993. Do communities di�er in health beha-
viours?. Journal of Clinical Epidemiology 46, 1141±1149.
Doll R., Peto R., 1981. The Causes of Cancer. Open
University Press, London.
Dunbar, G.C., Morgan, D.D.V., 1987. The changing pattern
of alcohol consumption in England and Wales 1978±1985.
British Medical Journal 295, 807±810.
Duncan, C., Jones, K., Moon, G., 1993. Do places matter?A
multilevel analysis of regional variations in health-related
behaviour in Britain. Social Science and Medicine 37, 725±
733.
ESRI, 1993. Arc/Info User Guide. Redlands, CA.
Gatrell, A.C., Dunn, C.E., Boyle, P.J., 1991. The relative uti-
lity of the Central Postcode Directory and Pinpoint
Address Code in applications of geographical information
systems. Environment and Planning A 23, 1447±1458.
Goldstein, H., 1991. Nonlinear multilevel models, with an ap-
plication to discrete response data. Biometrika 78, 45±51.
Goldstein, H., 1994. Improved estimation for logit and log-
linear multilevel models. Multilevel Modelling Newsletter 6,
2.
Goldstein, H., 1995. Multilevel Statistical Models. Edward
Arnold, London.
Goldstein, H., Spiegelhalter, D., 1996. League tables and their
limitations: statistical issues in comparisons of institutional
performance. Journal of the Royal Statistical Society A
159, 385±443.
Hart, C., Ecob, R., Davey Smith, G., 1997. People, places
and coronary heart disease risk factors: a multilevel analysis
C. Duncan et al. / Social Science & Medicine 48 (1999) 497±505504
of the Scottish Heart Health Study Archive. Social Science
and Medicine 45, 893±902.
Hauser, R.M., 1970. Context and consex: a cautionary tale.
American Journal of Sociology 75, 645±664.
Hilton, M.E., 1988. Regional diversity in United States drink-
ing practices. British Journal of Addiction 83, 519±532.
Humphreys, K., Carr-Hill, R., 1991. Area variations in health
outcomes: artefact or ecology. International Journal of
Epidemiology 20, 251±258.
Hunt, S.M., Martin, C.J., Platt, S., et al., 1988. Damp
Housing, Mould Growth and Health Status. Research Unit
on Health and Behavioral Change, Edinburgh.
Johnston, R.J., Pattie, C.J., Allsopp, J.G., 1988. A Nation
Dividing? Longman, London.
Jones, K., 1991. Using multilevel models for survey analysis.
Journal of the Market Research Society 35, 249±265.
Jones, K., Duncan, C., 1995. Individuals and their ecologies:
analyzing the geography of chronic illness within a multile-
vel modelling framework. Health and Place 1, 27±40.
Jones, K., Duncan, C., 1996. People and places: the multilevel
model as a general framework for the quantitative analysis
of geographical data. In: Longley, P., Batty, M. (Eds.),
Spatial Analysis: Modelling in a GIS Environment.
Longman, London.
Klein, H., Pittman, D.J., 1993. Regional di�erences in alcohol
consumption and drinkers' attitudes towards drinking.
American Journal of Drug and Alcohol Abuse 19, 523±538.
Kleinschmidt, I., Hills, M., Elliott, P., 1995. Smoking beha-
viour can be predicted by neighbourhood deprivation
measures. Journal of Epidemiology and Community Health
49 (S2), s72±s77.
Macintyre, S., 1996. What are spatial e�ects and how can we
measure them? In: Dale, A. (Ed.), Exploiting National
Survey and Census Data: the Role of Locality and Spatial
E�ects. CCSR Occasional Paper 12, University of
Manchester, Manchester.
Macintyre, S., MacIver, S., Sooman, A., 1993. Area, class
and health: should we be focusing on people or places?.
Journal of Social Policy 22, 213±234.
Marks, J.S., Hogelin, G.C., Gentry, E.M. et al, 1985. The
behavioral risk factor surveys. I. state-speci®c prevalence
estimates of behavioral risk factors. American Journal of
Preventative Medicine 1, 1±8.
Marmot, M.G., 1998. Improvement of social environment to
improve health. The Lancet 351, 57±60.
McKinlay, J.B., 1994. The promotion of health through
planned sociopolitical change: challenges for research and
policy. Social Science and Medicine 36, 109±117.
MMWR (Morbidity and Mortality Weekly Report), 1987.
Regional variation in smoking prevalence and cessation:
behavioral ®sk factor surveillance, 1986. Journal of the
American Medical Association 258, 3368±3370.
Oakley, A., 1994. Who cares for health: social relations, gen-
der and the public health. Journal of Epidemiology and
Community Health 48, 427±434.
Paterson, L., Goldstein, H., 1991. New statistical methods for
analysing social structures: an introduction to multilevel
models. British Educational Research Journal 17, 387±393.
Pierce, J.P., 1989. International comparisons in trends in ciga-
rette smoking. American Journal of Public Health 72, 152±
157.
Pugh, H., Power, C., Goldblatt, P. et al, 1991. Women's lung
cancer mortality, socio-economic status and changing
smoking patterns. Social Science and Medicine 32, 1105±
1110.
Rasbash, J., Woodhouse, G., 1995. MLn Command
Reference. Multilevel Models Project, Institute of
Education, University of London, London.
Rice, N., Leyland, A., 1996. Multilevel models: applications
to health data. Journal of Health Services Research and
Policy 1, 154±164.
Richards, R., 1989. Geographical distribution of cardiovascu-
lar ill-health and `risk factors': a small area analysis. Area
21, 306±307.
Rodriguez, G., Goldman, N., 1995. An assessment of esti-
mation procedures for multilevel models with binary re-
sponses. Journal of the Royal Statistical Society A 158, 73±
89.
Royal College of Physicians, 1983. Smoking or Health.
Pitman Educational, London.
Scott, A.J., Holt, D., 1982. The e�ects of two-stage sampling
on OLS methods. Journal of the American Statistical
Association 77, 848±854.
Skinner, C., Holt, D., Smith, T.F. (Eds.), 1989. The Analysis
of Complex Surveys. Wiley, New York.
Smyth, M., Brown, F., 1992. General Household Survey
Report 1990. HMSO, London.
Sooman, A., Macintyre, S., 1995. Health and perceptions of
the local environment in socially contrasting neighbour-
hoods in Glasgow. Health and Place 1, 15±26.
Syme, L., 1996. To prevent disease: the need for a new
approach. In: Blane, D., Brunner, E. (Eds.), Health and
Social Organisation: Towards a Health Policy for the 21st
Century. R. Wilkinson. London: Routledge.
Twigg, L., Moon, G., Jones, K., 1998. Predicting small area
health-related behaviour: a comparison of smoking and
drinking indicators. In: Earickson, R. and Schneider, D.
(Eds.), Proceedings of the Eighth International Medical
Geography Symposium. University of Maryland,
Baltimore.
Whichelow, M.J., Erzinclioglu, S.W., Cox, B.D., 1991. Some
regional variations in dietary patterns in a random sample
of British adults. European Journal of Clinical Nutrition
45, 253±262.
White, A., Nicolaas, G., Foster, K. et al., 1993. Health
Survey for England 1991. HMSO, London.
C. Duncan et al. / Social Science & Medicine 48 (1999) 497±505 505