Content uploaded by Kelvyn Jones
Author content
All content in this area was uploaded by Kelvyn Jones
Content may be subject to copyright.
HEALTH EDUCATION RESEARCH Vol.15 no.5 2000
Theory & Practice Pages 569–580
Consumed with worry: ‘unsafe’ alcohol consumption
and self-reported problem drinking in England
Liz Twigg, Graham Moon, Craig Duncan and Kelvyn Jones
Abstract cludes, however, that the two measures are
broadly similar in their relationship to social
and structural variables. Tenure provides anUsing data from the 1994 Health Survey for
England, logistic multivariate multilevel model- exception to this conclusion and indicates a
continuing need to take account of housingling techniques are used to investigate the simul-
taneous effect of individual demographic circumstances in developing an understanding
of drinking behaviour.characteristics and socio-structural factors on
self-reported problem drinking as revealed by
IntroductionCAGE scores and ‘unsafe’ levels of alcohol
consumption. Whilst the influence of key socio-
structural variables is broadly similar for both Whilst it is commonly accepted that low to moder-
ate levels of alcohol consumption may be beneficialunsafe alcohol consumption and high CAGE
scores, there are notable exceptions when results to health (Joint Working Group, 1995), long-term
high alcohol consumption has a negative impactare examined by tenure group: those in the
rented sector are more likely to be problem on health status (Anderson et al., 1993). The
Department of Health has published guidelinesdrinkers as revealed by CAGE, but less likely
to consume ‘unsafe’ amounts of alcohol. Both relating to ‘safe’ levels of consumption and these
currently stand at between 3 and 4 units a day fordimensions of drinking behaviour are influenced
by the consumption patterns of others in the men and between 2 and 3 for women (Interdepart-
mental Working Group, 1995). In addition, it ishousehold, with both likelihoods increasing as
the average consumption of others in the house- stated that drinking regularly 4 units or more per
day for men and 3 units per day or more forhold rises. After taking into account individual
compositional variables, the research indicates women will increase risks to health. Approximately
27% of men and 11% of women are consumingthat there is very little evidence for geographical
variation remaining in these two dimensions above these limits (Alcohol Concern, 1999). Health
of the Nation goals were set to reduce these to 18of drinking behaviour. It is found that the
proportion of the population whose drinking and 7%, respectively, for men and women by the
year 2005 (Department of Health, 1992). The 1999behaviour may be classed as (potentially) prob-
lematic via the CAGE responses is substantially White Paper on public health did not adjust these
goals (Secretary of State for Health, 1999).less than the proportion consuming above
recommended ‘safe’ levels. The research con- To help achieve success in reaching these targets,
it is useful for groups such as community alcohol
teams and the various voluntary bodies providing
alcohol-related services to know more about the
Institute for the Geography of Health, Department of
types of people who are likely to consume unsafe
Geography, University of Portsmouth, Buckingham
Building, Lion Terrace, Portsmouth PO1 3AS, UK amounts of alcohol, and where or in what type of
© Oxford University Press 2000. All rights reserved 569
L. Twigg et al.
area they may be located. There are, however, As Sutton and Godfrey (Sutton and Godfrey,
1995) note, data used in the analyses of consump-difficulties in the measurement and recording of
drinking behaviour. This paper addresses these tion patterns may be problematic, and may suffer
from measurement error and measurement bias.issues by examining information from the Health
Survey for England (HSE) on two measures relat- There are, for example, difficulties surrounding the
attempt to derive weekly estimates on the ‘usual’ing to problem drinking: alcohol consumption and
responses to a series of questions (known as the amount drunk by asking people to recall their
drinking behaviour over a 12 month period.‘CAGE questions’) which attempt to identify an
individual with (potentially) problematic drinking People’s patterns of drinking vary tremendously
from week to week, and individuals are oftenbehaviour.
The paper falls into three sections. The first unaware of what constitutes (say) a strong beer or
lager and may get confused over the amountsis concerned with the measurement of drinking
behaviour, and discusses consumption measures consumed. Furthermore, for various reasons, indi-
viduals may choose not to be truthful about theirand the CAGE questionnaire. The second section
outlines the methods employed in the present study. consumption patterns and underestimate their regu-
lar drinking level (Ehrens and Hedges, 1999).It provides brief descriptions of the data source
and the method of multivariate multilevel model- An alternative or validating measure that can be
used to assess levels of problematic drinking isling. The third section of the paper presents the
empirical results of the study. It compares and the CAGE score. This score is based on the
responses to a set of questions, known as thecontrasts the influence of individual demographic
and socio-structural factors in explaining high CAGE questions, which were originally developed
by Ewing and Rouse as a tool to help identifyCAGE scores and high alcohol consumption pat-
terns. It goes on to examine the relationships alcoholics in a hospital setting (Ewing and Rouse,
1970; Ewing, 1984). These questions are includedbetween individuals and the geographical contexts
in which they live with respect to the two dimen- in the 1994 HSE (Colhoun and Prescott-Clarke,
1996). Interviewees are asked to indicate a ‘yes’sions of drinking behaviour, and their covariance
across district health authorities (DHAs) and post- or ‘no’ response to six statements about drinking—
three on physical dependency and three on socialcode sectors.
attitudes. Work with CAGE has shown that a
positive response on two or more questions indi-Measuring drinking behaviour
cates a high likelihood of the presence of problem-
atic drinking (Mayfield et al., 1974).Patterns of alcohol consumption at the national
level are derived via government-sponsored sur- Several authors have evaluated the CAGE ques-
tions. Mayfield et al. (Mayfield et al., 1974) foundveys such as the HSE (Prescott-Clarke and Prim-
atesta, 1998). In the HSE, the respondent is asked that the questions correctly identified 87% of
alcoholic psychiatric patients. Masur and Monteiroto state how often they have consumed various
types of alcoholic beverage (e.g. shandy, beer, (Masur and Monteiro, 1983) found a 62% success
rate. Escobar et al. (Escobar et al., 1995) foundwine, spirits, etc.) over the last 12 months (e.g.
‘every day’, ‘once or twice a week’, ‘once or twice CAGE had the greatest efficacy and discriminating
power for diagnosing patients with alcoholism ina year’, etc.). The respondent is then asked to state
how much s/he usually drinks of this particular primary care when compared to other diagnostic
tests. A similar result was found when comparingtype of alcohol in any given day. This information
is used to generate an approximate weekly count CAGE questions with various chemical marker
tests (Girela et al., 1994). Lee and DeFrank (Leeof units of alcohol. Information from the HSE is
used to monitor progress towards national public and DeFrank, 1988) have found CAGE to be a
more successful indicator of self-reported drinkinghealth targets.
570
Consumed with worry
in men compared to women. They and Davidson tool in community settings both inside and outside
the USA where it was originally derived.
(Davidson, 1987) also levy some criticism at CAGE
in so far as time scales are not explicitly defined
Methods
in the questions and the same score or emphasis
can, for example, be allocated to a person who has
This paper considers the covariation of high CAGE
taken an ‘eye opener’ in the last few days and
scores and unsafe levels of alcohol consumption.
someone who took such a drink several years ago.
It also examines the joint relationship of both these
To this end, when comparing CAGE with another
outcome measures to socio-structural variables. It
screening test—the Michigan Alcohol Screening
has been widely illustrated that a number of vari-
Test (MAST)—it has been concluded that CAGE
ables, as well as age and gender, are important in
best identifies alcohol dependence in the previous
explaining alcohol consumption patterns at the
year (Watson et al., 1995).
individual level (Karvonen, 1995; Marmot, 1997).
The majority of work involving CAGE centres
Others have recognized the additional importance
on the identification of (potential) problem drinkers
of contextual influences in the form of drinking
in a clinical setting with relatively small numbers
patterns of other members of the family or house-
of individuals. There has only been a limited
hold (Rice et al., 1998; Sutton and Godfrey,
amount of work, which has systematically
1995). Our approach was therefore to compare and
described the socio-structural patterns of CAGE
contrast the effects of the main socio-structural
scores amongst the general population. Twigg et al.
variables of gender, age, social class, tenure and
(Twigg et al., 1998), for example, have compared
marital status on the likelihood of consuming
the overall proportion of problematic drinkers
unsafe amounts of alcohol and on the probability
in England and Que
´
bec, and provide a detailed
of scoring highly on CAGE, while incorporating
description of Que
´
becois people who score highly
the effect of household context by including a
on CAGE. One of the main reasons for the absence
household drinking index (Sutton and Godfrey,
of systematic community-level studies using
1995). Additionally, a dummy variable was used
CAGE has been that extensive national survey
to indicate whether the individual was from a
information has been unavailable until relatively
single person household.
recently. This shortcoming has been countered
Evidence also exists to suggest that geographical
by the inclusion of CAGE in major government
area of residence may influence consumption even
sponsored surveys. A second reason is that CAGE
after the compositional mix of an area has been
has traditionally been regarded as a tool to be used
taken into account (Blaxter, 1990; Williams and
in a clinical setting. However, Spencer et al.
Debakey, 1992; Duncan et al., 1993a; Karvonen,
(Spencer et al., 1987) have found CAGE to be
1995). This contextual influence may be perhaps
useful in screening for alcohol problems in the
best characterized by reference to the possible
general community and Colhoun et al. (Colhoun
impact of ‘cultural milieu’ on drinking behaviour.
et al., 1997) have used CAGE as a validation
To this end, the drinking behaviour of individuals
measure in an investigation of the collectivity of
can only fully be understood by reference to the
drinking in England. Dent et al. (Dent et al., 1995)
context in which the behaviour is learned and
compared CAGE responses to the Readiness to
performed on a day-to-day basis (Dorn, 1980).
Change questionnaire (RCQ) and argue that CAGE
While one such context is undoubtedly the house-
is a useful tool not only for identifying patients
hold, another is the area of a person’s residence.
who are problematic drinkers but also patients who
A popular manifestation of this latter effect would
are ready to make changes to address their drinking
be the reputed existence of ‘heavy-drinking’ envir-
problems. There is therefore a body of work that
onments. Macintyre et al. (Macintyre et al., 1993)
develop these arguments further within a broadersuggests that CAGE has utility as a measurement
571
L. Twigg et al.
Table I. High CAGE scores versus ‘unsafe’ alcohol Table II. Mean number of units consumed by CAGE score
consumption
No. of positive replies to Mean no. of units consumed by
the CAGE questions CAGE ScoreHigh alcohol consumption
Yes No Men Women
0 15.07 6.49High Cage Score Yes 588 (18.2%) 243 (2.3%)
No 2631 (81.8%) 10223 (97.6%) 1 31.7 16.7
2 40.91 20.91Total 3219 (100%) 10466 (100%)
3 46.33 21.68
Percentages of the column totals are given in brackets.
4 60.44 31.20
5 68.34 50.66
6 70.16 45.89
framework and question whether public health
should focus on people or places.
The influence of area or ecological context is of unsafe consumption using current Department
of Health definitions. However, as expected, mostinvestigated here by analysing data from the 1994
HSE in a multilevel model. The 1994 HSE is the of the ‘safe’ drinkers (98%) do not score highly
on CAGE. Table II shows a gender breakdown offourth wave of an annually repeated cross-sectional
survey in which interviewing takes place continu- the relationship between CAGE score and con-
sumption. This confirms a general positive trend,ously throughout the year. It uses a multi-stratified
design and is deemed to be a representative sample and the expected distinction between men and
women.of the English population (Colhoun and Prescott-
Clarke, 1996). A similar set of questions relating The major substantive research advance in this
paper concerns the simultaneous investigation ofto core topics is asked each year and questions on
alcohol consumption and drinking behaviour make both alcohol consumption patterns and high CAGE
scores in a multivariate multilevel model. Thisup part of this core. In total, approximately 15 800
individuals in the 1994 HSE supplied information approach provides a number of advantages over
the use of separate multilevel models for eachon their drinking behaviour, although only those
who stated that they drink ‘regularly’ were asked dimension of drinking behaviour (Duncan, 1997).
First, the relative influence of any one variable canthe CAGE questions. Both the alcohol consumption
variable and the results of the CAGE questions be assessed simultaneously for each dimension of
drinking behaviour. Much of the literature, forwere transformed into dichotomous variables.
Women consuming more than 14 units per week example, points to strong gender differences in
alcohol consumption (Blaxter, 1990; Duncan et al.,were classed as ‘unsafe’ drinkers; for men this
threshold was raised to 21 units. These definitions 1993a; Twigg et al., 2000), but do these also exist
for problem drinkers as revealed by CAGE? Indeed,of problem drinking were chosen to reflect tradi-
tional government guidelines regarding unit con- do gender differences occur across both dimensions
but with different strength and/or direction ofsumption and the literature on CAGE. All
individuals answering ‘yes’ to two or more of the influence? Second, not only can these questions
be explored more efficiently for both dimensionsCAGE questions were classed as high CAGE.
In the final analysis 13 685 individuals provided in a single multivariate model, the approach can
also allow for statistical testing of the differentialinformation on the two measures of drinking beha-
viour and a summary of the results is shown as a co-presence of compositional effects. Whilst vary-
ing effects for any one compositional variable couldcross-tabulation in Table I. Table I indicates that
the majority of ‘unsafe’ drinkers (82%) do not be obtained from separate models, the statistical
significance of any differences can only be gaugedscore highly on CAGE. This would suggest that
CAGE cannot be used as an indicator of levels when the models are run simultaneously. Third,
572
Consumed with worry
the multivariate multilevel approach allows for a categorical in nature and, for these, dummies were
simultaneous assessment of the importance of place
created that were modelled against a base category
with respect to both measures of alcohol related
denoting the modal characteristics for each vari-
behaviour. For example, the proportion of people
able. Age, which ranged from 16 to 97, was used
with CAGE problems may vary substantially from
as a continuous measure and centred on mean age.
place to place, but consumption may not. In the
The overall base category in the analysis was a
model reported in this paper, the measurement of
woman, aged 45, from social class I or II, who
both outcomes on a logit scale makes these pro-
was an owner occupier with a mortgage and was
cesses of comparison and significance testing rela-
married or cohabiting. The household drinking
tively straightforward. It is, however, possible to
index was also centred and a dummy variable
run mixed multivariate models with outcomes
indicating single person households was included
measured on different scales (Duncan, 1997).
in the model to allow investigation of the influence
Further, and perhaps more importantly, a major
of such households. Table IV summarizes the
advantage of modelling the two responses in a
variables used in the analysis.
single multilevel model is the ability to estimate
MLn software was used to carry out the
higher-level covariance terms. This joint covari-
multivariate multilevel models (Rasbash and
ance can illustrate the extent and manner in which
Woodhouse, 1995). A four-level model was
the two drinking behaviours covary across geo-
calibrated with level 1 representing the response
graphies. We would expect areas with high propor-
set (CAGE and units consumed; n ⫽ 27 370)
tions of CAGE problem drinkers to be
and level 2 the individuals (n ⫽ 13 685). Level
characterized by high levels of alcohol consump-
3 made use of the primary sampling unit (PSU)
tion but, through an analysis of residuals from the
codes (n ⫽ 712). These are based on postcode
higher levels, we can investigate those places or
sectors and whilst they remain un-named, they
types of place that do not fit this pattern. There
can still be used in the model because individuals
may, for example, be areas or types of area where
can be grouped according to PSU. Whilst
consumption is high, but the population is not
conclusions cannot be made about drinking
acknowledging problematic drinking patterns.
behaviour within identifiable PSUs, variation
Conversely, there may be areas where consumption
remaining at this level after taking account
is relatively low, but self-diagnosed problem
of individual composition can be determined.
drinking is declared to be relatively high.
Incorporating the PSU as a level in the analysis
A number of individual variables were used in
enables account to be taken of the clustering
the model to explain both dimensions of drinking
inherent in the survey sample design of the HSE.
behaviour. These were age, gender, social class,
At the same time, it also allows for an assessment
tenure, marital status and household drinking beha-
of possible small-scale ‘neighbourhood’ effects
viour. Table III provides the sample breakdown of
on drinking behaviour. PSUs are approximately
these variables and also indicates the levels of
the same size as electoral wards (8000–10 000
‘unsafe’ consumption and problematic CAGE
households). Other possibly more relevant small-
scores for each variable. The household drinking
scale area identifiers were not available. Level
index was calculated as described by Sutton and
4 is represented by named DHAs (n ⫽ 177)
Godfrey (Sutton and Godfrey, 1995) and summar-
and any variation remaining at this level can be
izes the drinking habits of other individuals within
estimated and residuals identified for named
the household. Where there is only one person in
geographical regions. DHAs were again included
the household, the value is set to the average for
for design reasons and also for their relevance
the index.
to policy and the planning of alcohol-related
Apart from age and the household drinking
index, all of the variables used in this paper were services.
573
L. Twigg et al.
Table III. Sample breakdown by unsafe drinking and CAGE scores
Variable No. in subgroup (percent of No. of ‘unsafe’ consumers in No. of CAGE-defined
group total in brackets) subgroup (percent of ‘problem’ drinkers (percent of
subgroup in brackets) subgroup in brackets)
Age
16–24 1787 (13.1%) 546 (30.6%) 157 (8.8%)
25–34 2844 (20.8%) 743 (26.1%) 228 (8.0%)
35–44 2545 (18.6%) 666 (26.2%) 191 (7.5%)
45–54 2169 (15.8%) 547 (25.2%) 133 (6.1%)
55–64 1793 (13.1%) 358 (20.0%) 76 (4.2%)
65–74 1614 (11.8%) 262 (16.2%) 33 (2.0%)
75⫹ 933 (6.8%) 97 (10.4%) 13 (1.4%)
Sex
male 6416 (46.9%) 2054 (32.0%) 538 (8.4%)
female 7269 (53.1%) 1165 (16.0%) 293 (4.0%)
Social class
I/II 3977 (29.1%) 1105 (27.8%) 273 (6.9%)
III non-manual 3332 (24.3%) 629 (18.9%) 153 (4.6%)
III manual 2695 (19.7%) 764 (28.3%) 184 (6.8%)
IV/V 3196 (23.4%) 636 (19.9%) 204 (6.4%)
class missing 485 (3.5%) 85 (17.5%) 17 (3.5%)
Marital status
married/cohabiting 8949 (65.4%) 1990 (22.2%) 477 (5.3%)
SWDS 4736 (34.6%) 1229 (26.0%) 354 (7.5%)
Tenure
owned (mortgage) 7064 (51.6%) 1807 (25.6%) 427 (6.0%)
owned (outright) 2985 (21.8%) 571 (19.1%) 103 (3.5%)
rent (LA/HA) 2264 (16.5%) 444 (19.6%) 166 (7.3%)
rent (private) 1237 (9.0%) 357 (28.9%) 129 (10.4%)
rent-free/missing 135 (1.0%) 40 (29.6%) 6 (4.4%)
Single-person household
no 10106 (73.8%) 2489 (24.6%) 635 (6.3%)
yes 3579 (26.2%) 730 (20.4%) 196 (5.5%)
Household drinking index
0.001–0.499 5579 (40.8%) 492 (8.8%) 182 (3.3%)
0.5–0.999 2444 (17.9%) 380 (15.5%) 117 (4.8%)
1.0–1.499 1553 (11.3%) 375 (24.1%) 87 (5.6%)
1.5–1.999 1195 (8.7%) 435 (36.4%) 91 (7.6%)
2.0–2.999 1203 (8.8%) 608 (50.5%) 104 (8.6%)
3.0–3.999 774 (5.7%) 409 (52.8%) 88 (11.4%)
4.0–4.999 346 |(2.5%) 178 (51.4%) 45 (13.0%)
5.0–39.0 591 (4.3%) 342 (57.9%) 117 (19.8%)
on the CAGE questions (CAGE) and on levels of
alcohol consumption (SAFE). Using a pseudo Z-Results and discussion
test, both of these logit estimates can be regarded
as significant as they are more than twice theirThe results are shown in Table V. The first row in
Table V represent the results for the constant term standard error. More sense of the values can,
however, be made by taking an anti-logit to obtainand show the estimated logit for the base category
574
Consumed with worry
Table IV. Variables used
Age centred on mean of 45 years
Gender male, female
Social class I/II, III non-manual, III manual, IV/V, class missing
Marital status married/cohabiting, single/widowed/divorced/separated
Tenure owned with mortgage, owned outright, housing association/local authority rented, private
rented, rent-free/missing
Household drinking index see text for an explanation of its calculation
Single-person household yes, no
Definitions underlined denote the base category or the ‘stereotypical’ respondent.
Table V. Composition results
Parameter Problematic drinking (CAGE) Unsafe alcohol consumption
(SAFE)
Estimate Standard error Estimate Standard error
Constant (CAGE, SAFE) 3.445* 0.1005 –1.829* 0.05777
Age as a linear term (AGECAGE, AGESAFE) –0.0232* 0.003929 –0.0105* 0.002048
Gender (i.e. male) (MALECAGE, MALESAFE) 0.923* 0.08802 1.111* 0.04883
Age⫻sex interaction (AGESEXCAGE, AGESEXSAFE) 0.001904 0.004682 –0.003939 0.00251
Single status (SINGCAGE, SINGSAFE) 0.2094* 0.08865 0.281* 0.05396
Social class IV/V (IV/VCAGE, IV/VSAFE) –0.1153 0.102 –0.3105* 0.06235
Social class III manual (3MCAGE, 3MSAFE) –0.179 0.1016 –0.06678 0.05941
Social class III non-manual (3NMCAGE, 3NMSAFE) –0.2629* 0.1087 –0.2724* 0.06209
Social class missing (CLMISSCAGE, CLMISSAFE) –1.145* 0.265 –0.7458* 0.1367
Own outright (OWNOUTCAGE, OWNOUTSAFE) –0.1279 0.1204 –0.02964 0.06286
Public renting (LAHACAGE, LAHASAFE) 0.3855* 0.1039 –0.1416* 0.06774
Private renting (PRIVCAGE, PRIVSAFE) 0.3434* 0.112 –0.03064 0.07617
Tenure missing/rent-free (TMISSCAGE, TMISSAFE) –0.3881 0.4223 0.1653 0.2073
Household drinking index (INDEXCAGE, INDEXSAFE) 0.265* 0.02442 0.5049* 0.02071
Single-person household (ONEHSCAGE, ONEHSSAFE) 0.0259 0.09907 0.004139 0.05941
Those values where the estimate is found to be significant using a pseudo Z-test are indicated with an asterisk. For those
variables where significance is achieved for both of the response variables (i.e. CAGE and unsafe levels of consumption), those
pairs that are statistically different from each other are shown in bold italics.
an estimate of the probability of these events to CAGE or units consumed. This effect of
age is significant for both the CAGE questionsoccurring. Thus the probability of a base category
respondent scoring highly on CAGE or being an (AGECAGE) and unsafe alcohol consumption
(AGESAFE), but the influence of age on the CAGEunsafe drinker is approximately 0.03 (i.e. a 3%
chance) and 0.14 (i.e. a 14% chance), respectively. variable is slightly stronger than its effect on unsafe
drinking behaviour. Although the magnitude of theThe remaining coefficients are interpreted with
respect to these constants and indicate differential difference between the CAGE and the consumption
coefficients is relatively small, further testing usingeffects for each variable allowing for other vari-
ables and the multilevel structure. For example, a contrast hypothesis testing procedure based on
the χ
2
statistic (Duncan et al., 1993b) reveals thatpeople above average age are less likely to be
classed as problematic drinkers, either by reference it is statistically significant (P 艋 0.05). This
575
L. Twigg et al.
suggests that age has a more moderating effect to reduce the probability for both dimensions of
problematic drinking. The relationship, however,on problematic drinking as revealed by CAGE
compared to unit consumption. Non-linear func- is not significant for the CAGE variable (IV/
VCAGE). Being in social class IV or V reducestions of age were explored but were found not to
have an effect on the model. the chances of consuming unsafe levels of alcohol
to approximately 11% for an otherwise stereotyp-The effect of being male on each of the responses
is both positive and statistically significant ical respondent (IV/VSAFE). The same pattern also
arises for the manual (3MCAGE and 3MSAFE) and(MALECAGE and MALESAFE). By adding these
terms to the constants and taking anti-logits, it can non-manual social class III groups (3NMCAGE
and 3NMSAFE). In the latter case, both dimensionsbe estimated that the probability for a male scoring
highly on CAGE or being an unsafe drinker rises of drinking behaviour have logit values that are
statistically significant. Further statistical testingto 7 and 33%, respectively. There is a slightly
higher gender effect with regard to levels of confirms that there is no significant difference in
magnitude between the two dimensions: beingconsumption but this difference is not statistically
significant. Age–gender interactions were also from a non-manual social class III background has
a similar effect on CAGE score and patterns ofincluded in the model (AGESEXCAGE and
AGESEXSAFE) but these were found not to be unsafe alcohol consumption. Taken together, these
findings challenge popular stereotypes of problemsignificant. CAGE and consumption thus appear
to perform similarly with regard to their association drinking as a characteristic of lower social class
individuals. They may well reflect hidden unmeas-with gender. Being male is clearly linked to both
high levels of consumption and to higher levels of ured relationships with disposable income, the
non-participation of people with serious drinkingself-assessed problem drinking. These results offer
individual-level confirmation of the ecologically problems in surveys such as the HSE or class-
related bias in estimating drinking. For those indi-based conclusions of the study by Colhoun et al.
(Colhoun et al., 1997) on regional prevalences viduals where class cannot be assigned (CLMISSC-
AGE and CLMISSAFE), a significant reduction inof heavy drinking and CAGE-defined problem
drinking. both the probability of scoring highly on CAGE
and in the probability of being categorized as anThe effect of marital status is statistically signi-
ficant, but its influence is not as strong as gender. unsafe drinker is shown. This appears particularly
influential for the CAGE variable, where the prob-Being single has a positive effect that is of similar
strength for both CAGE (SINGCAGE) and safe ability is reduced to approximately 1%. However,
statistical testing indicates no significant differenceconsumption (SINGSAFE). For a single woman
(in contrast to the typical individual, a married between the two logit values.
In terms of tenure, the logit values for owningwoman), the chance of scoring highly on CAGE
rises to 4 and 18% for unsafe consumption. The outright (OWNOUTCAGE and OWNOUTSAFE)
are not statistically significant, and it can beimplication here is that the presence of a spouse
or partner has a moderating effect that is similar assumed that there are no real differences between
this group and the base category (i.e. ownedfor both dimensions of drinking behaviour. There
are expected differentials between men and women through a mortgage). However, the values given
for local authority/housing association rentingin terms of the ‘protective’ effect of marriage or
cohabitation: for CAGE-defined problem drinking, (LAHACAGE and LAHASAFE) are significant
and interestingly the direction of influence is differ-marriage or cohabitation provides an additional
one percentage ‘protection’ for men, for the con- ent for the two dimensions of drinking behaviour.
A similar situation also occurs for the privatesumption measure there is a two percentage
point effect. rented sector where again the result is to reduce
the chances of being an unsafe consumer (PRIV-Lower social class (i.e. social class IV/V) tends
576
Consumed with worry
Table VI. Results: higher level variation
SAFE) but increase the probability of scoring
highly on CAGE (PRIVCAGE). Unlike social
Estimate Standard error
class, the missing category for tenure (TMISSC-
Level 4: DHAAGE and TMISSAFE), which also contains those
CAGE/CAGE 0.01404 0.0271
few individuals who live ‘rent-free’, is not signi-
SAFE/CAGE –0.00296 0.01348
ficant on either of the dimensions of drinking
SAFE/SAFE 0.03404 0.01309
behaviour.
Level 3: PSU
The findings concerning rented tenures are of
CAGE/CAGE 0.06432 0.05649
significance. Whilst being in the LAHA tenure
SAFE/CAGE 0.1004 0.02371
category reduces the probability of being an unsafe
SAFE/SAFE 0.01317 0.01958
drinker (as revealed by units consumed—LAHAS-
AFE) to approximately 12%, the chances of being
a problem drinker as revealed by CAGE is
increased to just over 4% (LAHACAGE). Statist- drinking index, the chances of consuming unsafe
amounts of alcohol for the otherwise typical indi-ical testing reveals that these within-tenure differ-
ences are unlikely to have occurred through chance vidual increase to 21%. The chances of a high
score on CAGE increase to just 4%. The effect ofprocesses. It would seem therefore that people
from rented tenures are less likely to report high single person households on the two dimensions
of drinking behaviour can be estimated using thelevels of consumption but more likely to be con-
cerned about their drinking habits. Explanations logit values for the dummy variables used to denote
such households (ONEHSCAGE and ONEHS-for this finding are likely to be complex and, in
the context of this paper, only speculations can be AFE). These are found to be positive for both
dimensions but neither is statistically significant.made. It may be that the relative poverty associated
with rented tenures leads to a lessened ability to Their influence on the response variables is
minimal.afford alcohol. This hypothesis is strengthened by
the reduction in the probability of being an unsafe Table VI indicates the estimates of the higher
level variances from the multivariate, multileveldrinker that is also experienced by people in lower
social classes. At the same time, it is possible that analysis with level 4 representing the 177 DHAs
and level 3 representing the 712 PSUs. It shouldpeople in rented tenures are more likely to under-
report their levels of consumption. Higher levels be noted here that the figures relate to the variances
and covariances of the model intercepts. Randomof CAGE-defined problem drinking may result
from the impact of differential patterns of drinking slopes were not allowed for in the analyses. The
covariance terms (SAFE/CAGE) measure theand consequential greater awareness of its problem-
atic aspects among people in rented tenures— degree to which the two dimensions of problem
drinking are related within the contexts of differentevidence suggests that binge drinking, for example,
is more common among people with characteristics levels. These terms cannot, however, be considered
in isolation. If there is to be a meaningful relation-associated with rented tenures (Moore et al., 1994).
The influence of household on drinking behavi- ship, in terms of covariation, then the two compon-
ent aspects of drinking behaviour should alsoour is summarized by reference to the drinking
habits of other members of a respondent’s house- exhibit a significant degree of variation within the
relevant level. Table VI indicates that there ishold. The derived drinking index (INDEXCAGE
and INDEXSAFE) is influential on both dimen- significant intercept variation, albeit small, across
DHAs for the SAFE response variable but not forsions of drinking. Both logit values are significant
and positive but the value for unsafe/safe levels of the CAGE variable. There cannot therefore be any
meaningful covariance at this level notwithstandingconsumption (INDEXSAFE) is almost twice that
of the CAGE response. For a unit increase in the the fact that the reported negative covariance
577
L. Twigg et al.
is, in any case, not statistically significant. This suggested that age, sex, marital status, class and
household drinking patterns all exert similar influ-suggests that, at the DHA level there is little
unexplained contextual variation. The composi- ences on both dimensions of drinking behaviour.
Consumption levels are reduced and CAGE scorestional mix of an area (i.e. the characteristics of its
individuals) is of most importance in explaining lowered for older people, women and married (or
cohabiting) people. Levels rise and CAGE scoresgeographical variation in both CAGE scores and
levels of unsafe consumption. This finding under- increase for those in the higher social classes, a
trend that challenges notions that people fromlines the importance of going beneath the ecolo-
gical analysis reported in the earlier study by higher classes may be less likely to define heavy
drinking as problematic. As expected, those indi-Colhoun et al. (Colhoun et al., 1997). Moreover,
though there is a very small negative covariance viduals living in households where other members
are consuming high levels of alcohol are alsoat this level, the lack of significance of that
covariance and one of its component parts (CAGE/ themselves more likely to be heavy drinkers and
classed as a (potentially) problematic drinkerCAGE) indicates that there is no evidence that
there are DHAs where the population is on average, according to the CAGE questionnaire. This overall
picture of similarity between the two dimensionsdrinking unsafe amounts of alcohol but not scoring
highly on CAGE. At the PSU level there appears of drinking behaviour does not hold for tenure.
Rented tenures appear to exert different influencesto be significant positive covariance suggesting
that, at that level, areas with high CAGE scores on each dimension such that people in rented
accommodation are more likely to have high CAGEalso have high levels of consumption after taking
account of the compositional mix. This finding is, scores and less likely to drink to unsafe levels.
Given the spatial concentration of these tenurehowever, undermined by a lack of significance in
the two component parts of the covariance. In groups in the UK, this finding has implications for
the delivery of alcohol-related services and healthsummary, Table VI suggests that there is very little
ecological context or ‘real’ geography to problem promotion.
It is acknowledged that the measurement ofdrinking as revealed by CAGE scores or units
consumed. Instead, variation is determined by drinking behaviour is a problematic and contested
field. The analysis presented in this paper hasthe characteristics of the people living in those
geographical areas. employed established definitions of problem
drinking drawn from government policy and pub-
Conclusions lished literature. It must be noted that these defini-
tions are conservative and may exaggerate the
number of unsafe drinkers with serious problemsThe nature and extent of drinking behaviour is an
important consideration in the formulation of pub- of alcohol misuse. The focus of the paper is not,
however, on the utility of these survey instrumentslic health strategy. This paper has explored the
dimensions of drinking behaviour through an but rather with an investigation of the covariation
of the two aspects of drinking behaviour whichinnovative multivariate multilevel analysis of two
distinct measures of that behaviour. By undertaking they purport to measure. In this regard the paper
has provided a substantive contribution indicatingsimultaneous analysis of ‘unsafe’ drinking as meas-
ured by number of units consumed and self- the particular nature of that covariation taking
account of the influence of contextual circum-reported problem drinking as reflected by high
CAGE scores, the paper has been able to distin- stances.
Key analytical gains from the simultaneousguish between quantitative aspects of drinking
behaviour and more qualitative perceptions of multivariate multilevel approach are evident in the
above findings. Only by considering consumptionapproaches to drinking.
The analysis presented within the paper has levels and CAGE scores within the same model
578
Consumed with worry
models in geographical research. In Westert, G. P. and
can we directly compare the two behaviours and
Verhoeff, R. N. (eds), Places and People: Multilevel
assess the relative influence of any one variable
Modelling in Geographical Research. Netherlands
Geographical Studies 227. The Royal Dutch Geographical
on both dimensions of drinking behaviour at the
Society, Utrecht, pp. 100–119.
same time. This is particularly evident in the
Duncan, C., Jones, K. and Moon, G. (1993a) Do places matter?
application of simultaneous significance testing
A multilevel analysis of regional variations in health-related
behaviour in Britain. Social Science and Medicine, 37,
across all variables for each dimension of drinking
725–733.
behaviour. In the reported analysis, application of
Duncan, C., Jones, K. and Moon, G. (1993b) Blood pressure,
this procedure resulted in a recognition that age
age and gender. In Woodhouse, G. (ed.), A Guide to ML3
for New Users, 2nd edn. Institute of Education, University
and household levels of drinking have statistically
of London, London, pp. 55–82.
distinct relationships with each measure of problem
Ehrens, B. and Hedges, B. (1999) Alcohol consumption. In
drinking—albeit in the same direction. This proced-
Prescott-Clarke, P. and Primatesta, P. (eds), Health Survey
for England: The Health of Young People 1995–1997. A
ure also confirmed the existence of a differential
Survey carried out on behalf of The Department of Health
influence exerted by the rented tenures on each
by the Joint Health Surveys Unit of Social and Community
aspect of drinking behaviour. Analysis of higher-
Planning Research (SCPR) and the Department of
Epidemiology and Public Health, University College London.
level intercept variation and covariation was also
Volume 1: Findings. The Stationery Office, London, chap. 7.
facilitated by the multivariate approach and
Escobar, F., Espi, F. and Canteras, M. (1995) Diagnostic tests
revealed that there was little evidence for regional
for alcoholism in primary health care: compared efficacy of
different instruments. Drug and Alcohol Dependence, 40,
cultures of drinking. Once individual character-
151–158.
istics have been taken into account, there is also
Ewing, J. A. (1984) Detecting alcoholism: the CAGE
no significant evidence that there are places where
questionnaire. Journal of the American Medical Association,
252, 1905–1907.
CAGE scores and consumption do not co-vary as
Ewing, J. A. and Rouse, B. A. (1970) Identifying the hidden
expected.
alcoholic. Paper presented at the 29th International Congress
on Alcohol and Drug Dependence, Sydney, Australia.
Girela, E., Villanueva, E., Hernandez-Cueto, C. and Lunda, J.
References
D. (1994) Comparison of the CAGE questionnaire versus
some biochemical markers in the diagnosis of alcoholism.
Alcohol Concern (1999) Information for Purchasers: The
Alcohol and Alcoholism, 29, 337–343.
Scale of the Problem. (http:/www.alcoholconcern.org.uk/
Interdepartmental Working Group (1995) Sensible drinking.
purchasers.prchinfo.htm).
Department of Health, London.
Anderson, P., Cremona, A., Paton, A., Turner, C. and Wallace,
Joint Working Group of Royal College Of Physicians, Royal
P. (1993) The risk of alcohol. Addiction, 88, 1493–1508.
College of Psychiatrists, Royal College of General
Blaxter, M. (1990) Health and Lifestyles. Tavistock/
Practitioners (1995) Alcohol and the Heart in Perspective—
Routledge, London.
Sensible Limits Reaffirmed. Royal Colleges, London, pp.
Colhoun, H. and Prescott-Clarke, P. (1996) Health Survey
1–36.
for England 1994, Volume II: Survey Methodology and
Karvonen, S. (1995) Regional differences in drinking among
Documentation. HMSO, London.
Finnish adolescents. Addiction, 90,57–64.
Colhoun, H., Ben-Shlomo, Y., Dong, W., Bost, L. and Marmot,
Lee, D. J. and DeFrank, R. S. (1988) Interrelationships among
M. (1997) Ecological analysis of collectivity of alcohol
self-reported alcohol intake, physiological indices and
consumption in England: importance of average drinker.
alcoholism screening measures. Journal of Studies on
British Medical Journal, 314, 1164–1168.
Alcohol, 49, 532–537.
Davidson, R. (1987) Assessment of the alcohol dependence
Macintyre, S., MacIver, S. and Sooman, A. (1993) Area, class
syndrome: a review of self-report screening questionnaires.
and health: should we be focusing on places or people?
British Journal of Clinical Psychology, 26, 243–255.
Journal of Social Policy, 22, 213–234.
Dent, T. H. S., Shepherd, R. M., Alexander, G. J. M. and
Marmot, M. (1997) Inequality, deprivation and alcohol use.
London, M. (1995) Do CAGE scores predict readiness to
Addiction, 92 (Suppl. 1), S13–S20.
reduce alcohol consumption in medical in-patients? Alcohol
Masur, J. and Monteiro, M. G. (1983) Validation of the CAGE
and Alcoholism, 30, 577–580.
alcoholism-screening test in a Brazilian psychiatric inpatient
Department of Health (1992) The Health of the Nation: A
hospital setting. Brazilian Journal of Medical and Biological
Strategy for Health in England. HMSO, London.
Research, 16, 215–218.
Dorn, N. (1980) Alcohol in teenage cultures: a materialist
Mayfield, D., McLeod, G. and Hall, P. (1974) The CAGE
approach to youth cultures, drinking and health education.
questionnaire: validation of a new alcoholism-screening
Health Education Journal, 3,67–73.
instrument. American Journal of Psychiatry, 131, 1121–1123.
Moore, L., Smith, C. and Catford, J. (1994) Binge drinking—Duncan, C. (1997) Applying mixed multivariate multilevel
579
L. Twigg et al.
Sutton, M. and Godfrey, C. (1995) A grouped data regressionprevalence, patterns and policy. Health Education Research,
9, 497–505. approach to estimating economic and social influences on
individual drinking behaviour. Health Economics, 4, 237–Prescott-Clarke, P. and Primatesta, P. (eds) (1998) Health Survey
for England 1996: A Survey carried out on behalf of The 247.
Twigg, L., Moon, G. and Pampalon, R. (1998) Person, placeDepartment of Health by the Joint Health Surveys Unit of
Social and Community Planning Research (SCPR) and the and the geographies of problem drinking: a multilevel
comparison of England and Que
´
bec. Allocation des resources,Department of Epidemiology and Public Health, University
College London. The Stationery Office, London. ge
´
ographie des soins, Actes du Ve
`
me Colloque Ge
´
ographie
et socio-e
´
conomie de la sante
´
. CREDES, Paris, pp 63–72.Rasbash, J. and Woodhouse, G. (1995) MLn Command
Reference. Multilevel Models Project, Institute of Education, Twigg, L., Moon, G. and Jones, K. (2000) Predicting small-
area health-related behaviour: a comparison of smoking andUniversity of London, London.
Rice, N., Carr-Hill, R., Dixon, P. and Sutton, M. (1998) The drinking indicators. Social Science and Medicine, in press.
Watson, C. G., Debra, E., Fox, K. L. and Ewing, J. W.influence of households on drinking behaviour: a multilevel
analysis. Social Science and Medicine, 46, 971–979. (1995) Comparative concurrent validities of five alcoholism
measures in a psychiatric hospital. Journal of ClinicalSecretary of State for Health (1999) Saving Lives: Our Healthier
Nation (Cm 4386). The Stationery Office, London. Psychology, 51, 676–684.
Williams, G. and Debakey, S. (1992) Changes in levels ofSpencer, J., Bartu, A. and Harrison-Stewart, A. (1987)
Observations on community screening for alcohol problems: alcohol consumption: United States, 1983–1988. British
Journal of Addiction, 87, 643–648.a pilot project to assess the feasibility of identifying heavy
drinkers in a community setting. Alcohol and Alcoholism,
Received on October 19, 1999; accepted on March 5, 200022,65–69.
580
