Content uploaded by Kelvyn Jones
Author content
All content in this area was uploaded by Kelvyn Jones on Nov 24, 2017
Content may be subject to copyright.
Sot. Sci. Med. Vol. 31, No. 6, pp. 725-133,
1993 0277-9536/93
$6.00 + 0.00
Printed in Great Britain. All rights reserved Copyright 0 1993 Pergamon Press Ltd
DO PLACES MATTER? A MULTI-LEVEL ANALYSIS OF
REGIONAL VARIATIONS IN HEALTH-RELATED
BEHAVIOUR IN BRITAIN
CRAIG DUNCAN,’ KELVYN JONES’ and GRAHAM MOON’
’ Department of Geography, University of Portsmouth, Buckingham Building, Lion Terrace, Portsmouth
PO1 3HE, England and *School of Social and Historical Studies, University of Portsmouth, Milldam,
Burnaby Road, Portsmouth PO1 3AS, England
Abstract-A number of commentators have argued that there is a distinctive geography of health-related
behaviour. Behaviour has to be understood not only in terms of individual characteristics, but also in
relation to local cultures. Places matter, and the context in which behaviour takes place is crucial for
understanding and policy. Previous empirical research has been unable to operationalize these ideas and
take simultaneous account of both individual compositional and aggregate contextual factors. The present
paper addresses this shortcoming through a multi-level analysis of smoking and drinking behaviours
recorded in a large-scale national survey. It suggests that place, expressed as regional differences, may be
less important than previously implied.
Key words-multi-level model, lifestyle behaviour, smoking, drinking, geography, place differences
INTRODUCTION
Most medical geographic research has used geo-
graphic space merely as a framework in which data
can be ordered [l]. This extremely limited role for
geography in the structuring of mortality, morbidity
and health care activity rests on the simplistic use of
space as “an organising framework for recognising
regular associations” [2, p. 3991. These regular associ-
ations are usually assessed by statistical models which
are assumed to apply uniformly across study areas
and in which people are often reduced to ecological
indicators. A conceptualization of people as agents
and, what is more, of people as agents influenced by
the different place situations in which they find
themselves, is missing. This perspective is restrictive;
as Jones and Moon argue:
Space and locality have a crucial role to play in explanation
because the various processes concerned can be manifested
in different ways at different places [2, p. 2661.
This restrictiveness is particularly evident in studies
purporting to examine health-related behaviours such
as smoking, drinking, diet, exercise, sexuality and
drug misuse. These “lifestyles” might be expected to
result from both individual predisposition and geo-
graphically-based, cultural influences. Within recent
years such studies have grown in number, being
based, for the most part, on “lifestyle” questionnaires
conducted throughout the United Kingdom at a
variety of scales. Although such surveys constitute an
important secondary data source, and have revealed
much about aggregate trends in health-related be-
haviour, they have been little used in explorations of
the difference that place makes in the structuration of
lifestyle behaviour.
This paper aims to address this shortcoming. It
presents an assessment of the role of place in structur-
ing smoking and drinking behaviour in the United
Kingdom and begins with a section on “lifestyle”
surveys. Attention then turns to an outline of a
quantitative approach to the establishment of the
reality and nature of geographic variations in health-
related behaviour. The following section summarizes
the results of this investigation and places them
within the context of existing knowledge about smok-
ing and drinking behaviour. The implications of these
results are outlined, and a short conclusion then
considers the limitations of the findings.
HEALTH AND LIFESTYLE
Throughout the 1980s in the United Kingdom
researchers consistently documented the existence of
continuing, and in some cases widening, health in-
equalities between social classes [3]. Considerable
academic and political debate raged over the various
explanations proposed for these inequalities. The
balance of medical and social scientific opinion ar-
gued that the key factor was material structural
disadvantage leading either directly, or indirectly
through constraining effects, to a greater probability
of poorer health and earlier death amongst people in
lower status occupations. This opinion was not how-
ever unanimous and alternative explanations focused
on the roles of social selection [4] and health-related
behaviours.
The behavioural or “lifestyle” explanation argued
that independent and autonomous behaviour on the
part of an individual generated morbidity or mor-
tality [3, p. 1 lo]. Individuals, or groups of individuals
125
726 CRAIG DUNCAN et al.
behaving in a systematically similar fashion, were
held to be individually responsible for the health status). A three-stage sampling design was followed
in which parliamentary constituencies within each
outcomes of inappropriate behaviour. In its raw standard region were randomly selected from three
unconsidered form the explanation implied that in- bandings of population density with probabilities
equalities in health were the fault of the sufferer’s proportional to the constituency populations. Two
lifestyle or deviant culturally-guided behaviour. This electoral wards were selected from each of the 198
could be identified through lifestyle surveys. Targeted constituencies on a similar basis and individuals were
health education messages of the “look after your- then selected randomly from the wards. In total the
self’ variety would then place the onus on the survey recorded data on 9003 individuals in 396
individual to choose to correct “behavioural deficits”. different locations across mainland Britain.
With its central thesis of individualism and freedom A number of workers have carried out secondary
of choice, the explanation found considerable politi- analyses of the HALS data [lo] and it forms the basis
cal favour with the Conservative Governments of the of the most important publication to date on British
1980s; its most trenchant manifestation was in the lifestyle behaviour-Mildred Blaxter’s Health and
utterings of Edwina Currie, sometime junior minister Lifestyles [1 11. Accounting for geographical vari-
of health, regarding the dietary habits of inhabitants ations in “lifestyle” was one of Blaxter’s primary
of Northern England. objectives, her main concern being to:
Such a basic approach to health-related behaviour
has been questioned by a number of commentators.
The Health Divide questions the extent to which
behaviour can be abstracted from its social context
[3]. Others have noted that:
distinguish between the importance of those elements of
lifestyle which are typical of socio-economic status and
those elements which are truly regional or associated with
the external living environment [l 1, p. 791.
it would run counter to the evidence to assume that people’s
patterns of smoking, drinking, eating and sexual activity are
determined by individual choices that are unaffected by
social, economic or legislative factors [S, p. 751.
From a geographical perspective, it seems inher-
ently likely that location could be added to these
influences.
Away from the political and academic debates
surrounding health inequality, health-related be-
haviours remained, throughout the 198Os, the oper-
ational subject-matter of health education and, to a
lesser extent, community medicine. Lifestyle surveys
provided guidance for initiatives targeted on those
sections of the community exhibiting risk behaviours;
they also provided epidemiological assessments of the
prevalence of such behaviour. More recently the rise
of the “new public health” has questioned the effec-
tiveness of targeted health education [6] but lifestyle
surveys have gained the support of health service
management as vehicles for the health needs assess-
ments central to effective purchasing in the reformed
National Health Service [7]. The attempt in 1991 to
begin the development of a health policy for England
saw further emphasis being placed on lifestyle and
behaviourally-oriented health promotion strategies
PI.
This concern can be rephrased to indicate the need to
establish the extent to which geographical variations
in health-related behaviour are a consequence of the
varying social compositions of areas, or of the more
subjective cultural context: the spirit of place. Places
with high levels of smoking, for example, may simply
be composed of more people with individual charac-
teristics indicating a predisposition to smoking. Alter-
natively, all people in that place, regardless of their
individual, personal characteristics may be affected
by contextual, ecological factors (e.g. a regional
culture that encourages smoking). Both sets of factors
may operate-or only one.
Like other writers [12], both before and after
HALS, Blaxter supports the idea of significant con-
textual effects. In relation to smoking she writes:
. class is related to smoking in different ways in different
types of area [I 1, p. 1171.
and for alcohol consumption she concludes:
the relationship of social class to drinking habits depends
very much on the environment [l 1, p. 1191.
Many individual health authorities have now un-
dertaken health and lifestyle surveys and, in 1984/5,
a national survey was carried out by the University
of Cambridge School of Clinical Medicine. The
“Health and Lifestyle Survey” (HALS) focused on:
“lifestyles, behaviours and circumstances relating to
the physical and mental health of the population” [9,
p. 11. It collected detailed data on behaviours together
with additional health information (e.g. attitudes and
physiological measures) and individual background
characteristics (e.g. income, occupation, age, marital
Blaxter thus emphasizes that people’s health-re-
lated behaviours are very much influenced by the
places in which they live. The overriding message is
that geography matters enormously in lifestyle be-
haviour; indeed Blaxter states that this is one of the
more important conclusions of her analysis [1 1, p.
2361.
Unfortunately, this message is based upon a highly
simplistic and problematic statistical analysis. Blax-
ter’s preferred methodology is to calculate standard-
ized ratios for different subgroups. In effect, the
multilevel nature of both the problem and the data
(people in areas) is lost as she is forced to work at a
single aggregate level, grouping together qualitatively
heterogenous individuals to ensure statistically re-
liable rates based on large numbers of respondents.
Regional variations in health-related behaviour in Britain 72-l
This results in crude geographical and statistical
assemblages of the form that Sayer disparagingly
calls chaotic conceptions and taxonomic collectives
[ 131. Thus, people become manual and non-manual
classes, while place becomes North vs the South. This
approach aggregates individuals and confounds the
compositional with the contextual.
METHODOLOGICAL ISSUES
HALS was based on a multi-stage sampling design
which generates an inherently hierarchical data set.
Individuals, within wards, within constituencies,
within regions were randomly selected. This data
structure is ideally suited to a multi-level approach
which distinguishes contextual and compositional
differences, does not rely on crude geographical ag-
gregations, and maintains the original complex
nature of the data [14]. In this way, multi-level
modelling differs from both Blaxter’s tabular risk
analysis approach and other aggregate strategies in
acknowledging that it is individuals, not groups or
places that behave, thereby avoiding the ecological
fallacy. Moreover, and unlike-individual-level logistic
modelling [15], it recognizes that individuals behave
in context, thus avoiding the atomistic fallacy [16].
The fundamentals of multi-level modelling as ap-
plied to medical geographic problems have been
comprehensively covered elsewhere [17] and will not
be rehearsed here. This present paper focuses on a
three-level analysis of the HALS data offering simul-
taneous consideration of i individuals, nested in j
wards in k regions [18]. The wards and the regions
provide levels at which the contextual impacts on
compositional level-one effects can be assessed. While
the wards are used in the analysis, it is results at the
regional level that will be most fully explored.
Two multi-level models are central to the analysis:
Y,,,=Box,+(ek+~jk++Eilk) (1)
(Ok + pjk +&(/k) t2)
Equation (1) is a simple “null” three-level, random-
intercepts model in which variation in the response
variable Y is “explained” by the single fixed intercept
term &, the national average, and three random
terms associated with this intercept which reflect the
remaining variation at the individual level ciikr the
ward level pjk, and the region level Ok. This variation
can be summarised in three variance terms; ai, a:,
and u: which estimate respectively the between-indi-
vidual, between-ward and between-region variation.
This null model is expanded in equation (2) to include
m fixed, national effects (the /Is) associated with m
“predictor” variables (the Xs) for individuals. The
actual equation (2) used in this paper employed eight
level-one predictors: age, sex, social class, school-
leaving age, employment status, housing tenure and
marital status which are specified as a set of 23
dummy, indicator variables. In this model, the two
higher level random terms now assess the ward and
regional differences after taking into account individ-
ual characteristics which may be influencing health-
related behaviour. If the multi-level models reveal no
significant nor substantial remaining variation at the
second or third levels, this would imply, within the
limitations of model specification, that there are no
contextual influences on health-related behaviour at
these broad geographical levels. If this were to be the
case, there would be no distinctive geography of
“lifestyle”, only a sociology. Conversely, if the
higher-level variation increases following the inser-
tion of level-one variables, then it would be possible
to conclude that geographical difference is more
marked than the apparent sociology would suggest
]191.
This paper focuses on two health-related be-
haviours: drinking and smoking. The data on drink-
ing consist of units of alcohol consumed in a week.
This is a continuous variable and can therefore be
modelled using the usual linear link function [20]. The
smoking data, however, consist of a binary response
variable. Individuals either are or are not current
regular smokers. Such data could be modelled by
treating the response as if it were the probability of
smoking. However, this practice creates three prob-
lems: non-sensical values, inappropriate functional
form, and variance heterogeneity. The appropriate
solution is to use a non-linear form which can be
linearized for estimation by taking a logit transform-
ation and specifying a binomial error term [21]. All
models were calibrated using the software package
VARCL3 [22].
In all the fitted models the predictor, X, variables
were coded to indicate departures from the most
common variable characteristic [23]. The effect of this
strategy is that the intercept term in the model (/IO)
effectively represents the stereotypical individual sur-
vey respondent. This enables straightforward in-
terpretation of the model output. The exact profile of
the stereotypical individuals is described more fully in
the relevant following section.
THE GEOGRAPHY OF HEALTH-RELATED BEHAVIOUR
This section details the results obtained from the
multi-level modelling of the data recorded by HALS
for smoking and drinking. It is separated into two
sub-sections, one for each behaviour.
Analysis of smoking behaviour
Smoking contributes to deaths from coronary
heart disease and many forms of cancer; it is the
major cause of death from cancer of the lung [24]. Its
health consequences are not only direct: “passive
smoking” is thought to be responsible for some 20%
of lung cancer cases [25]. These causes of death are
particularly significant in the United Kingdom con-
728 CRAIG DUNCAN et al.
text [26] where some 30% of women and 33% of men
smoke [8]. Cigarette smoking behaviour first became
common in the earlier half of this century; while
women developed the habit more recently than men.
Evidence from routine statistical sources such as the
General Household Survey suggests that there has
been a considerable decline in the number of smokers
over the past quarter-century but that young people,
particularly women, have been less affected by this
trend; they also indicate a pronounced social class
relationship. Smoking rates remain amenable to re-
duction and targets for age and sex-specific re-
ductions have been proposed [8, p. 671.
A traditional geographical analysis of smoking
behaviour would aggregate the number of individuals
found smoking in each area, calculate the rate for
each area and map or tabulate the results [27]. Table
1 shows that the Health and Lifestyle Survey
recorded considerable variation in smoking be-
haviour around the country. The lowest level of
smoking was recorded in Devon and Cornwall
(76.2% of the sample did not smoke regularly), whilst
the highest was found in Inner London (only 56.6%
did not smoke regularly). This variation may relate to
contextual, ecological factors or it may be an artefact
created by different regional compositions of different
types of individuals: Devon and Cornwall’s popu-
lation could be composed of significantly more people
who are unlikely to smoke due to age, class and other
important consumption cleavages.
Table 2 (column A) gives the summary statistics for
the null model of smoking behaviour [equation (l)].
The intercept, Do, gives a nation-wide, average esti-
mate of an individual smoking. When transformed
from its logit value, this represents a 66% chance that
any individual is a non-smoker. This probability
Table I. Aggregate percentage rates of people NOT
smoking by region based on the 1984/1985 Health &
Lifestyle Survey
Average percentage
Region pop. NOT smoking
Strathcylde 58.5% (20)
E. Central Scotland 62.5% (17)
Rural Scotland 63.0% (16)
Rural North 66.5% (1 I)
Industrial North East 63.8% (13)
Merseyside 57.4% (21)
Gtr Manchester 58.7% (19)
Rest of North West 65.0% (12)
West Yorkshire 63.4% (15)
South Yorkshire 70.1% (8)
Rural Wales 67.3% (10)
Industrial South Wales 59.0% (18)
W. Mids Conurbation 63.5% (14)
Rest of West Midlands 68.4% (9)
E. Midlands 70.8% (7)
E. Anglia 71.8% (4)
Devon and Cornwall 76.2% (I)
Wessex 71.5% (5)
Inner London 56.6% (22)
Outer London 71.0% (6)
Outer Metropolitan 72.4% (3)
Outer South East 73.8% (2)
Note: values in parentheses represent rank position,
1 = least smoke, 22 = most smoke.
Table 2. Multi-level estimates for models of smoking behaviour
(A) (B)
Fixed erects
Level I
Intercept
sex
Male
Social class
Other non-manual
Employers/managers
Semi-skilled
Professional
Unskilled
Age
16-24 yr
25-39 yr
6&54 yr
65-74 yr
75-97 yr
Age leave school
<14yr
14yr
16yr
17yr
18yr
19+ yr
Employment status
Unemployed
Housing status
Local Authority Renter
Other Renter
Marital status
Single
Widowed
Divorced
Random effects variance
Level I
Intercept, 0:
Level 2
Intercept, 0:
Lode13
-0.67 -0.65
0.14 (2.91)
-0.37 (-5.20)
-0.14 (-1.98)
0.04 (0.55)
-0.73 (-5.15)
0.18 (1.68)
0.14 (1.40)
0.09 (1.41)
-0.36 (-3.17)
-0.83 (-6.36)
-1.60 (-9.32)
-0.17 (-1.15)
-0.12 (-1.56)
-0.33 (-4.79)
-0.31 (-3.16)
-0.71 (-6.65)
-0.47 (-2.12)
0.52 (4.89)
0.69 (11.8)
0.56 (6.46)
0.004 (0.06)
0.004 (0.04)
0.48 (4.97)
1 .oo I .oo
0.11 (9.89) 0.05 (5.27)
Intercept, 0: 0.04 (4.50) 0.01 (2.57)
Note: estimates represent logit values; figures in parentheses rep
resent ratio of estimates to standard error.
value does not remain constant over the country and
the values for the random effects-a:, 0: and &-
show to which level the remaining variability can be
apportioned. This reveals that, whilst the majority of
variation in smoking (86.5%) occurs at level 1 (that
is between individuals), ca 10% and 4% occurs at
levels 2 and 3 respectively. This higher-level variation
can be assessed for significance by calculating the
ratio of the estimate to its standard error. If the ratio
is in excess of plus or minus two, the estimate is
considered significantly different from zero. As shown
in Table 2, both values are significant and the differ-
ences at both higher levels are more than could be
expected from sampling fluctuations.
It would seem, therefore, that there is some degree
of significant ward and regional variation in smoking
behaviour which does not originate in sampling
fluctuations. The 22, level-three (region) differences
from equation (1) are plotted in their logit form in
Fig. 1. These differences measure the reduction or
increase in comparison to the national logit of being
a smoker in each region. When transformed, they
indicate that the probability of an individual being a
non-smoker ranges from 59% in Inner London to
71% in the Outer South-East.
Regional variations in health-related behaviour in Britain 129
The null model of equation (1) therefore suggests
the tentative conclusion that there is a geography of
smoking, but the higher level variation may be an
artefact of compositional factor-ertain wards or
regions may have a preponderance of individuals
with characteristics associated with particular smok-
ing behaviours. This hypothesis was tested using
equation (2). The results are shown in Table 2,
column B. The stereotypical individual is an em-
ployed woman, aged 4@59 yr, married and living in
an owned house. She left school at 15 and is now in
the skilled manual social class category. The fixed
estimates in the model reveal that the chances of an
individual not smoking decrease significantly should
they be male, divorced, of a lower social class,
unemployed, or a Local Authority tenant. At the
same time those chances would increase significantly
were they older and of the professional class.
The inclusion of these compositional fixed terms
reduces the variance of the random intercepts at the
two higher, ward and regional, levels. The variance
that remains at these levels quantitatively measures
contextual effects on smoking; geographical vari-
ations in composition have been “controlled” by the
inclusion of the individual attributes. Only 4.5% of
the variance can be attributed to the ward level and
the regional variance has been reduced to c 1%.
Although both these values retain statistical signifi-
cance, it would appear that a large amount of the
geographical variation in smoking behaviour can be
attributed to compositional variations in the individ-
ual predisposition to smoke rather than to any
higher-level contextual geographical effect.
The regional geography of smoking which remains
after controlling for individual compositional factors
is illustrated by the plot for equation (2) in Fig. 1. The
plot reveals a generalized North-South gradient in
Least smoke
Strathclyde I-
smoking behaviour. Northern regions tend to have
higher levels of smokers than their Southern counter-
parts once social composition is controlled. It is
possible to use these estimates of region effects in
conjunction with the fixed level-one coefficients to
generate estimates of the probability of individuals
in particular localities being non-smokers: for
example, a professional woman in the Outer South
East has an 8 1% chance of not being a smoker whilst
this reduces by 3% should the same woman be in
Merseyside.
Taking the plots for equations (1) and (2) together,
Fig. 1 also illustrates the changes in regional variation
before and after composition is taken into account. It
appears that regional variation collapses once com-
positional factors are included. This is strikingly
conveyed by the convergence of most lines around the
value 0 (which represents /IO, the overall national
estimate). There are two exceptions: in Rural Wales
people do not smoke who would be expected to and
in the Rest of the North West the reverse is true.
Overall Fig. 1 confirms the limited role that ecological
factors play in regional smoking variations. There is
some regional variation-but it is relatively small and
to some extent originates in compositional effects.
When transformed to probabilities, the regional
differences after controlling for composition indicate
a variation of only a +3% variation around the
national average 65% chance that the stereotypical
individual is a non-smoker. The regional differences
in the probability of not smoking are clearly small in
relation to individual socio-demographic differences.
Analysis of drinking behaviour
Alcohol misuse is associated with cirrhosis of the
liver, digestive cancers, road traffic accidents, high
blood pressure and psycho-social problems; it has
Most smoke
E. Cent. Scotlind -
Rural Scotland -
Rural North -
Industrial N.E. -
Merseyside -
Gtr. Manchester -
Rest of North West -
West Yorks -
South Yorks -
Rural Wales -
Industrial S. Wales -
West Midlands Con. -
Rest of West Mids -
East Midlands -
East Anglia -
Devon and Cornwall -
Wessex -
Inner London -
Outer London -
Outer Metropolitan r
Outer South East L---l---.-.....
I I ’ I I I
-0.3 -0.2 -0.1 0 0.1 0.2 0.3 0
Residual
Fig. 1. Model comparison-smoking behaviour.
730 CRAIG DUNCAN et al.
Table 3. Aggregate rates of units of alcohol consumed
per person per week by region based on the 1984/1985
Health & Lifestyle Survey
Average units consumed
Region per person per wk
Strathcylde 12.87 (15)
E. Central Scotland 12.69 (14)
Rural Scotland 12.43 (13)
Rural North 11.01 (6)
Industrial North East 13.90 (18)
Merseyside 14.40 (21)
Gtr Manchester 13.30 (17)
Rest of North West 10.39 (5)
West Yorkshire 14.47 (22)
South Yorkshire 14.00 (19)
Rural Wales 9.63 (3)
Industrial South Wales 13.16 (16)
W. Mids Conurbation 14.30 (20)
Rest of West Midlands 10.04 (4)
E. Midlands 11.86 (11)
E. Anglia 8.56 (I)
Devon and Cornwall 1 I .06 (7)
Wessex 11.71 (IO)
Inner London 12.00 (12)
Outer London 1 I .46 (9)
Outer Metropolitan 11.22 (8)
Outer South East 8.79 (2)
Note: values in parentheses represent rank position,
1 = least drink, 22 = most drink.
been estimated to be implicated in at least 25,000
premature deaths in Britain per year [28]. The re-
search evidence suggests that alcohol is only damag-
ing when used in quantity. This is commonly assessed
by reference to “units’‘-a standard half-pint of beer,
measure of spirits or glass of wine-with “safe” limits
being set at 14 units for women and 21 for men [29].
The General Household Survey suggests that heavy
drinking is commonest among manual workers and
young single men and The Health of the Nation
recommends a target of less than 1 in 6 men and 1 in
18 women drinking over sensible limits by the year
2000 [8, p. 701.
As with smoking, a traditional analysis of drinking
would be based upon crude averages for geographical
areas [27]. Regional level averages for the Health
and Lifestyle Survey are given in Table 3 and
display apparent variation. The average number of
units of alcohol consumed in a week by an individual
ranges from 8.56 in East Anglia to 14.47 in West
Yorkshire.
A constant-only, “null” multi-level model
[equation (l)] was fitted and the results are summar-
ized in Table 4, column A. The estimated intercept,
&, , represents the average number of units consumed
over the entire country and is estimated at 11.87. This
value does not remain constant across all regions and
wards and the variation in consumption can be
decomposed to each of these higher levels. When this
is done the majority of the variation (98%) occurs at
level 1, between individuals. However, ca 1% can be
attributed to each of the higher levels and this
variation is in fact significant when the ratio test is
conducted. The region level residuals for drinking
from equation (1) are plotted in Fig. 2. On the basis
of these values, the average units consumed in a week
varies from 9.58 in the Outer South East to 13.33 in
the West Midlands Conurbation.
The results for the random-intercepts, equation (2)
model, taking account of compositional individual-
level factors, are given in Table 4, column B. In this
instance, the stereotypical individual changes as only
a subset of the 9003 individuals surveyed responded
to the questions concerning alcohol. The reduced
dataset of 6211 persons produces a stereotype who is
an employed, married, 25-39 yr old man, who left
school at 15, owns his own house and is in a skilled
manual occupation. With the addition of fixed esti-
mates, the random variation at the ward and region
level is further reduced, although both estimated
variances (cr: and ai) remain significant [30].
Important differences between the individual
effects on drinking and those for smoking can be
noted. Men drink more as well as smoke more but the
gender differential in relation to alcohol consumption
is much more significant. With smoking there is an
expected class gradient linking low social status to
smoking. For drinking this is reversed: the pro-
fessional and employer/manager social categories
consumed more alcohol than any of the lower social
classes (plus 0.95 and 1.12 units respectively). This
Table 4. Multi-level estimates for models of drinking behaviour
(A) (8)
Fixed effects
Level 1
Intercept
sex
Female
Social class
Other non-manual
Employers/managers
Semi-skilled
Professional
Unskilled
Age
1624 yr
4&59 yr
60-64 yr
65-74 yr
75-97 yr
Age leave school
<14yr
14yr
16yr
17yr
18yr
19+ yr
Employment mmo
Unemployed
Housing status
Local Authority Renter
Other Renter
Marital status
Single
Widowed
Divorced
Random effects variance
Level I
Intercept, 0:
Level 2
Intercept, ui
Level 3 3
11.87 18.41
-12.08 (-31.5)
-1.32 (-2.37)
1.12 (1.99)
- 0.82 ( I .40) -
0.94 (1.08)
-0.01 (-0.01)
0.14 (0.20)
-1.08 (-2.20)
-3.30 (-3.38)
-3.15 (-2.78)
- 5.59 (-4.00)
-2.55 (-2.06)
-0.75 (-1.19)
-1.19 (-2.24)
-0.67 (-0.91)
-0.62 (-0.85)
- 2.39 (- I .48)
1.75 (2.03)
1.42 (2.86)
0.50 (0.71)
3.67 (5.81)
1.85 (2.04)
3.45 (4.34)
253 210
2.53 (3.63) 1.71 (3.04)
Intercept, 0; 1.82 (4.05) 1.46 (4.04)
Note: estimates represent units of alcohol; figures in parentheses
represent ratio of estimates to standard error
Regional variations in health-related behaviour in Britain
Strathclyde
E. Cent. Scotland
Rural Scotland
Rural North
Industrial N.E.
Merseyside
Gtr. Manchester
Rest of North West
West Yorks
South Yorks
Rural Wales
Industrial S. Wales
West Midlands Con.
Rest of West Mids
East Midlands
East Anglia
Devon and Cornwall
Wessex
Inner London
Outer London
Outer Metropolitan
Outer South East -3
Least drink Most drink
m Equation (1)
0 Equation (2)
-2 -1 0 1
Residual
Fig. 2. Model comparison-drinking behaviour.
731
finding is contrary to the aggregate research evidence
(see above) but is marginal and not statistically
significant when the standard error ratio is calculated.
Overall, it would appear that whilst smoking be-
haviour is most closely related to social class and
tenure, drinking is determined more by gender differ-
ences and marital status. Both behaviours diminish
significantly with age.
The region residuals produced by the equation (2)
alcohol model are also shown in Fig. 2. They rep-
resent units of alcohol above and below the overall
national average, /I,,, for each individual region. A
North-South gradient is again visible. As with the
smoking analysis, the equation (1) and equation (2)
values for drinking can also be compared in Fig. 2,
giving a graphical representation of regional vari-
ation before and after regional composition is taken
into account. The plot for alcohol differs substan-
tially from that for smoking. The regional residuals
for drinking do not collapse around /I0 once compo-
sition is included. Rather, there are many cases of
absolute residual change (i.e. shifts away from the
average &). There is heavier drinking in West York-
shire, the Industrial North East, Merseyside, Greater
Manchester and the East Midlands than would be
expected on the basis of socio-demographic compo-
sition. In contrast, Inner London, East Central Scot-
land and the Outer Metropolitan area all return
unexpectedly low figures for alcohol consumption
given their demographic and social composition.
There is thus more support for greater ecological,
contextual variation at the region level in drinking
behaviour than smoking. However, this effect is still
small when judged against individual-level differ-
ences. The regional alcohol consumption range, after
controlling for composition, varies by only f 10%
from the overall national average of 18.41 units for
the stereotypical individual. The variation within a
region is much more substantial than between regions
[311.
Implications
On the basis of the multi-level analyses presented
above, it would appear that the major determinants
of geographical variation in two key lifestyle be-
haviours are individual-level factors. The direction
and strength of these individual-level effects is sub-
stantially in accordance with established research
knowledge. Suggestions of an important role for
contextual, geographical effect is, however, limited,
especially in relation to smoking behaviour. Places-
be they wards or regions-have little independent
effect. In both cases there is much greater variations
in behaviour between individuals than between wards
and regions even after taking account of individual
demographic and social characteristics. Most impor-
tantly, geographical contextual variables specified at
these higher levels cannot possibly account for this
individual-level variation, for they will be effectively
constant at this lower level.
It would appear that Blaxter’s conclusions about
the importance of a contextual “area” effect are
over-stated [ 111. There are geographical variations
but only as a consequence of differing place compo-
sition. Similarly, contentions that people of similar
age and occupation “tend to smoke and drink in a
manner strongly influenced by where they live” [32]
are undermined once compositional factors are prop-
erly and completely taken into account. That is not
to say, of course that other geographies or “settings”
such as the work-place, are not of importance.
For health policy, the implication of the analyses
presented in this paper is a questioning of the tra-
ditional belief that health care purchasers need to
732 CRAIG DUNCAN et al.
conduct specific, local lifestyle surveys and respond to
“unhealthy lifestyles” on an area-basis. The findings
presented here strongly challenge the wisdom of such
a dual approach. First, the limited role of contextual-
ity allows the imputation of national survey results to
local situations using a multi-level model in combi-
nation with key local data. This strategy has been
found to produce an accurate assessment of lifestyle
behaviour and has considerable cost-saving potential
[33]. Second, the importance of composition implies
that health promotion in relation to smoking and
drinking should not focus exclusively on area-based
preventive strategies.
CONCLUSIONS
The analyses presented here suggest that assumed
contextual geographical differences in lifestyle are not
substantial. This finding should not, however, be
taken as a denial of the geographical. First, it is based
on relatively unsophisticated models. It is likely that
individuals learn, experience and display health-re-
lated behaviours in particular social contexts; these
social contexts are likely to be manifested in geo-
graphical settings. Second, the level-l variables used
in this paper, whilst seemingly justified, could be
challenged for their adequacy, although the range of
available variables was obviously constrained by the
questionnaire adopted by the Health and Lifestyle
Survey research team. Third, the hierarchical defi-
nition of the levels could be criticized as an inappro-
priately formalistic and mechanistic attempt to
capture the cultural geography of lifestyle. The levels
reflect the data collection process and a convenient
regional classification, it may be argued, not local or
regional culture [34]. Fourth, given the novelty of
multi-level modelling, certain technical problem areas
remain imperfectly understood, notably the binomial
assumption of the level-l variance in models with a
binary response [17]. Finally, it should, of course, be
emphasized that a multi-level approach retains many
of the limitations of more traditional quantitative
medical geography; in this research, lifestyle be-
haviour is modelled more powerfully than traditional
techniques allow, yet it is still analysed in a rather
crude and mechnical manner lacking in the ability to
offer any insight into what causes people to smoke or
drink. Moreover, while there may be little quantitat-
ive difference, the qualitative nature of the behaviour
may be culturally specific. Nevertheless, within its
limitations, multilevel analysis can contribute to the
development of a place-sensitive medical geography
and the debunking of crude regional stereotypes of
health-related behaviour.
Acknowledgements-The authors would like to acknowl-
edge the extremely useful comments of two anonymous
referees and participants at the Ftfth International Sym-
posium in Medical Geography, Charlotte, NC, U.S.A.,
August, 1992.
2.
3.
4.
5.
6.
8.
9.
10.
11.
12.
13.
14.
15.
16.
REFERENCES
Exceptions include the ethnographies of Cornwell J.
Hard-earned Lives. London, Tavistock, 1984; Donovan
J. We Don’t Buy Sickness; it just comes. London,
Gower, 1986; Evles J. Sense of Place. Silverbrook.
Warrington, 1985; and Moon G.“Conceptions of space
and community in British health policy. Sot. Sci. Med.
30, 165-171, 1990.
Jones K. and Moon G. Health, Disease and Society.
RKP, London, 1987.
Townsend P., Davidson N. and Whitehead M. Inequal-
ities in Health. Pelican, London, 1988.
Illsley R. Occupational class, selection, and the pro-
duction of inequalities in health. Ortlv J. Sot. Affairs 2.
151-165, 1986: _ I “_ I
Smith A. and Jacobson B. The Nation’s Health. Kings
Fund, London, 1988.
Ashton J. and Seymour H. The New Public Health.
Open University Press, London, 1988.
Health Education Authority, Office of Population Cen-
suses and Surveys. Health and Lifestyle Surveys:
Towards a Common Approach. HEA, London, 1990.
HMSO. The Health of the Nation. HMSO, London,
1991.
Cox B. et al. The Health and Lifestyle Survey: Prelimi-
nary Report. Health Promotion Research Trust,
London, 1987.
See for example Humphreys K. and Carr-Hill R. Area1
variations in health outcomes: artifact or ecology. ht.
J. Epidemiol. 20, 1-8, 1991; Humphreys K. and Carr-
Hill R. Health and lifestyle: is there a cultural effect on
health behaviour? Dept. of Social Statistics, University
of Southampton, mimeo, 1991. The latter uses multilevel
modelling to examine variations in smoking, drinking
and diet, but unlike the present paper, it concentrates on
ward, and not regional differences.
Blaxter M. Health and Lifestyles. Tavistock/Routledge,
London, 1990.
See for example Shaper A. et al. British regional heart
study: cardiovascular risk factors in middle aged men in
24 towns. Bt. Med. J. 283, 179-186, 1981; Fox A.,
Goldblatt P. and Jones D. Social class mortality differ-
entials: artefact, selection or life circumstances? In Class
and Health (Edited by Wilkinson R.). Tavistock,
London; Fox A., Jones D. and Goldblatt P. Approaches
to studying the effect of socio-economic circumstances
on geographic differences in mortality in England and
Wales. Br. Med. Bull. 40, 309-314, 1984; Hart N.
Inequalities in health: the individual versus the environ-
ment. J. Roy. Statistical Society A 149, 228-246, 1986;
see also [lo].
Sayer A. Method in Social Science: A Realist Approach.
Hutchinson, London, 1984.
Goldstein H. Multi-level Models in Educational and
Social Research. Griffin. London. 1987.
Jones K. and Moon G: Predicting local variations in
health-related behaviour. Health Information Research
Service Working Paper 7. Portsmouth Polytechnic,
Portsmouth, 1992; Bucquet D. and Curtis S. Socio-
demographic variation in perceived illness and the use
of vrimarv care. Sot. Sri. Med. 23, 737-744, 1986:
Balarajan k., Yuen P. and Machin D..Deprivation and
general practitioner workload. Br. Med. J. 304,
529-534, 1992.
In terms of probabilistic inference, the use of single-level
individual models with hierarchically-structured data
sets can be anticipated to find significant relationships
where none exist. See Skinner C., Hold D. and Smith T.
(Eds) Analysis of Data from Complex Surveys. Wiley,
Chichester, 1989. Multi-level modelling allows for intra-
class correlation and automatically “adjusts” standard
errors.
Regional variations in health-related behaviour in Britain 733
17. See Jones K. Multi-level Models for Geographical Re- 27.
search. Geobooks. Norwich. 1991: Jones K. and Moon
I8.
19.
20.
21.
22.
23.
24.
25.
26.
G. A multi-level’ approach to immunisation uptake.
Area 22,264-271, 1990; Jones K. and Moon G. Re-as-
sessing immunisation uptake as a performance measure 28.
in general practice. Br. Med. J. 303, 28-31, 1991; Jones
K., Moon G. and Clegg A. Ecological and individual 29.
effects in childhood immunisation uptake: a multi-level
approach. Sot. Sci. Med. 33, 501-508, 1991. 30.
The usual division of the United Kingdom into 11
Standard Regions is too coarse for the purposes of this
paper and for the Healfh and Lfestyle Survey. Accord-
ingly, the constituencies (and hence) the wards are
grouped into the 22 regions as used by The Economist.
Essentially this grouping represents a subdivision of the
standard regions into metropolitan and non-metropoli-
tan areas.
Nelder J. and Wedderburn R. Generalised linear 31.
models. J. Roy. Stat. Society A 135, 370-384, 1972.
Jones K. and Moon G. Re-assessing immunization
uptake as a performance measure in general practice.
Br. Med. J. 303, 28-31, 1991.
Longford N. VARCL: interactive software for variance
components analysis. Professional Statistician 5, 28-32,
1986.
For an example of this outcome, see Jones K. and
Bullen N. A multilevel analysis of the variations in
domestic property prices: southern England 1980-1987.
Urban Studies forthcoming.
For an example of this approach to smoking and
drinking see Balarajan R. and Yuen P. British smoking
All the models fitted in this paper are random-intercept
and drinking habits: regional variations. Communify
Med. 8, 1311137, 1986. _
ones; that is they are based on the assumption that
Royal College of Physicians. The Medical Consequences
of Alcohol Abuse. Tavistock, London, 1987.
there is an overall difference between ward and
Royal College of Psychiatrists. Alcohol: Our Favourite
Drug. Tavistock, London, 1986.
regions. It is intended in future work to fit ‘random-
It will be noted that the variances from the two
models imply that the geography of drinking may, to
slope’ models in which, for example, the relationships
a small extent, be regional, while that for smoking
may be local, i.e. ward-based. Further research is
between smoking and class are allowed to vary
currently addressing this issue and additionally consid-
ering multivariate situations in which both behaviours
between areas. See [16] and [17] for examples of such
are examined simultaneously, and models which
allow the effects of individual attributes to vary at
models.
the individual level. Such developments are reviewed
in Jones K. Using multilevel models for survey analy-
sis. In Survey and Statistical Computing (Edited by
Westlake A. et al.), pp. 23 l-242. Elsevier, Amsterdam,
1992.
Jones K. Multi-level Models for Geographical Research. 32.
Geobooks, Norwich, 199 1.
Doll R. and Peto R. The Causes of Cancer. OUP,
London, 198 1. 33.
Wald N. and Namchahal K. Does breathing other
people’s tobacco smoke cause lung cancer? Br. Med. J.
293, 1217-1221, 1986.
Cummins R. ef al. Smoking and drinking by middle-
aged men: the importance of town of residence. Br.
Med. J. 2.83, 1497-1502, 1981.
Jones K. and Moon G. Predicting local variations in
health-related behaviour. Health Information Research
Service Working Paver 7. Portsmouh Polvtechnic.
Portsmouth, 1992. _
World Health Organisation. Controlling the Smoking 34, Further research is addressing this issue by developing
Epidemic. WHO Technical Report 636, WHO, Geneva, models which specify contextual variables at levels 2
1979. and 3.
