Tobacco Documents Online

~Per~mnon PII: S0277-9536(96)00431-5 Soe. Sci. Med. Vol. 45, No. 6, pp. 893-902, 1997 1997 Elsevier Scienc~ Ltd. All fights reserved Printed in Great Britain 0277-9536/97 $1%00 + 0.00 PEOPLE, PLACES AND CORONARY HEART DISEASE RISK FACTORS: A MULTILEVEL ANALYSIS OF THE SCOTTISH HEART HEALTH STUDY ARCHIVE CAROLE HART)* RUSSELL ECOW and GEORGE DAVEY SMITH~ 'Department of Public Health, University of Glasgow, 2 Lilybank Gardens, Glasgow Gl2 8RZ, U.K., :MRC Medical Sociology Unit, 6 Lilybank Gardens, Glasgow GI2 8RZ, U.K and 'Department of Social Medicine, University of Bristol, Canynge Hall, Whiteladies Road, Bristol BS8 2PR, U.K. Abstract--The Scottish Heart Health Study (SHHS), which recruited 5123 men and 5236 women between 1984 and 1986, was set up in part to investigate geographical variation in coronary heart dis- ease in Scotland. Multilevel models are particularly appropriate for such hierarchical data, in which the individuals in the study can be reprc~nted by the lower level and the districts in which they live by the higher level. Multilevel models are 15r~ented for four coronary heart diseas~ risk faetors~diastolic blood pressure, eholesterel, alcohol consumption (defined both as units of alcohol consumed per wcek and as being a non-drinker) and smoking, for men and women separately. Significant district level var- iance was found for three out of the four variables studied, after controlling for socioeconomic and other variables considered at the level of the individual. These were for diastolic blood pressure, choles- terol and alcohol. Although the large majority of the variance was present at the individual level the existence of sig- nificant variance at the district level is ¢viden~ that places may have a role in the distribution of coron- ary heart disease risk. Health policy aimed at reducing coronary heart disease should therefore consider the characteristics of places-as well as individuals. ~) 1997 Elsevier Sciene~ Ltd Key words--multilevel model, diastolic blood pressure, cholesterol, drinking, smoking, coronary heart disease INTRODUCTION Individual choice of behaviours which arc condu- cive to good health, such as healthy eating, is con- sidered to be the key factor which can lead to improvements in the nation's health (Scottish Office, 1992; Department of Health, 1992). However, health-related bchaviours, such as smok- ing, drinking and diet, may reflect both factors which arc characteristics of the individual (for example age, gender and socioeconomic position) and geographically-based cultural influences (Duncan et al., 1993). In addition to possible cul- tural differences between places, aspects of the area where individuals live.may help or hinder attempts to improve their health (Maeintyre et al., 1993). Examples include availability of sports facilities and shops selling healthy food at affordable prices. It is therefore important to consider both individuals and the areas in which they live (Jones and Moon, 1993; Duncan et al., 1996). *Author for correspondence: West of Scotland Cancer Surveillance Unit, Ruehill Hospital, Bilsland Drive, Glasgow G20 9NB, U.K. This paper explrres this issue by using the tech- nique of multilevel modelling, which considers both area and individual effects simultaneously (Goldstein, 1995). Data from the Scottish Heart Health Study (SHHS)-are used to investigate any additional effects of area, over and above the effects of factok's eharaeterised at the level of the individ- ual, for four coronary heart disease, risk factors-- diastolic blood pressure, cholesterol, alcohol con- sumption and smoking. Separate models are pro- duced for men and women. METHODS The SHHS is a study of coronary heart disease risk factors and lifestyle in Scotland. Twenty-two local government districts were chosen out of a total of 56 in mainland Scotland. A target total of 450 people aged 40-59 years from each district were randomly selected from general practices in the dis- triets, except for Edinburgh and North Glasgow where the target was 800. The achieved sample was 5123 men and 5236 women. The study was con- dueted between 1984 and 1986. Participants com- pleted a questionnaire and attended a sercening centre where physical measurements were taken. 893 This article is tbr individual use only and may not be thrther reproduced or stored electronically without written permission from the copyright holder.

894 Carole Har~ ~! oL Methods have previously been described in derail (Smith et eL, 1987; Tunstall-Pedoe et aI., 1989; Smith eta/., 1989). Multilevel models arc appropriate for data of this hierarchical type, as they allow for independent ran- dom effects at two or more levels (Goldstein, 1995). In this case we have two levels, the lower level (level I) consisting of the individuals and the higher level (level 2) of the districts in which they live. In the most straightforward models these are the var- iance components in the model. Multilevel models are particularly appropriate for surveys using multi stage sampling schemes such as that of the SHHS, though such sampling is not necessary for their use. The models also allow for explanatory variables at each level. The models presented here include only explanatory variables characterised at the individual level. The following variables relating to individuals were included in the models: age (centred about the mean of 49.6 years for men and 49.4 years for women), years of education, social class (coded as 1 for those with non manual occupations and 0 for those with manual occupations according to the Registrar General's classification), housing tenure (coded as I if an owner occupier and 0 if not) and, for the models with diastolic blood pressure and cholesterol as independent variables, body mass index (in kg/m2). For the diastolic blood pressure model only, the additional variable "medication for high blood pressure" (coded as I for yes and 0 for no) was included. For women who classed themselves as house- wives, social class was taken as that Of their hus- bands, otherwise the social class based on their own occupation was used. At the screening examination, two blood pressure readings were taken with the patient seated, using a Hawksley random zero sphygmomanometer. Measurements were made at the first and the fifth Korotkoff sounds (Smith et aL, 1990). The two dia- stolic blood pressure readings were averaged for the modelling analysis (in mmHg). A blood sample was also taken at the screening examination and total serum cholesterol (in mmol/l) was estimated by en- zymatic means (Smith et aL, 1989). A record of the amount of alcohol consumed on each day of the previous week was included in the questionnaire. This was converted to units of alco- hol assuming I unit per g/ass of wine, I un/t per half pint of beer and I unit per glass of spirits. Alcohol was modelled in two different ways. Firstly, the total units per week was used as the measure of alcohol. Secondly, alcohol consumption was defined as being a non-drinker, in line with cur- rent thinking on the protective effect of moderate alcohol consumption on coronary heart disease risk (Rimm et al., 1996). Smoking was defined as currently being a regular or occasional cigarette smoker from the question- naire responses. Two forms of models are used, depending on whether the dependent variable is continuous or discrete. For continuous variables (diastolic blood pressure, cholesterol and alcohol consumption in units per week), the models are of the form y#=a/+bx~+e#, where i represents the individual and j the district. y# is the response variable of interest (e.g. dia- stolic blood pressure). For simplicity, only one explanatory variable (xa) is shown, though more may be included. The quantities a and b are the fixed parameters of the model. The random part of the model consists of uj at the district level and e¢ at the individual level. It is assumed that uj and e# are normally distributed, independent between levels and have zero expected values. For discrete (0,I) variables (smoking and alcohol as non-drinking), a logistic regression model is used, given by y9 = exp(f# + uj)/(l + exp(J~j + u/)) + e# where~ denotes the fixed part of the model. The e# are assumed to have a binomial distribution. The multilevel modelling software package MLn was used (Woodbous~, 1995). In the case of discrete dependent variables, the model was converted into a form suitable for use by the software using a macro supplied as part of the software and the MQL option was used (Rodriguez and Goldman, 1995). The modelling procedure began with a baseline model using age only as an explanatory variable. Each variable was added separately and checked for significance by calculating if the estimat~ were greater than twice its standard error. Then the sig- nificant variables were included together. The sig- nificance of the d/strict level variance (at the 5% level) was assessed by a likelihood ratio test of the model against a model with district level variance constrained to zero. Standard errors are also shown against the estimates of the random effects at dis- trice level, but due to the positive skewness of the distribution of the variance, it is possible that a value which is less than two standard errors can reach statistical significance at the 5% level. For discrete dependent variables, the variance at the individual levd was set to !, corresponding to the assumption of no extra-binomial variation, and this assumption was tested in each logistic re- gression model. No evidence was found for extra- binomial variation. Evidence of extra-binomial vari- ation could be due to omission of an important level I ~xplanatory variable or omission of a higher level in the model (Woodhouse, 1995). i I ! I ! ! I I [ [

I People, places and CHD risk factors 895 Table I. Estimates from multilevel models of diastolic blood pressure (in mmHs) for men and women from the SHHS Men (n - 5026) Women (~ - 5152) Baseline Final Baseline Final Fixed effects ~onstant 84.5 (0.42) 64.2 (I.26) 81. I (0.42) 66.4 (0.96) age 0.17 (0.03) 0.11 (0.03) 0.32 (0.03) 0.21 (0.03) body ma. index 0.77 (0.05) 0.58 (0.03) medication for high blood pressure 6.79 (0.63) 6.03 (0.58) housing tenure -0.56 (0.33) --1.09 (0.32) Random effects level 1 131.1 (2.62) 120.3 (2.40) 126.4 (2.50) 114.8 (2.27) level 2 3.30 (1.17) 3.05 (1.08) 3.36 (I.18) 2.85 (I.01) Standard e~ror~ in parentheses. Level l--indlvidual level. Level 2--distrint level. In all models, individuals were excluded if they had missing data for any of the variables included in the final model. The maximum number of miss- ins cases was 796 for the cholesterol ,,model for women. RESULTS (a) Diastolic blood pressure Men. Variables not significant when added separ- ately to the model were years of education and social class. The baseline model with age only and the final model with other explanatory variables added are shown in Table 1. The variable housing tenure which was significant when added separately to the model became non significant when all were added. The model demonstrates that diastolic blood pressure increases with increasing age and body mass index, an.d is 6.79 mmHg higher for men on medication for high blood pressure. There was a small and significant district level variance in the baseline model Although the ma- jority of the variance occurred at the individual level, 2.5% occurred at the district level. When the explanatory variables were added, both the individ- ual and district level variances decreased but were still significant. The proportion of total variance explained at the district level was the same as in the baseline model. Women. The variables years of education and social class were not significant and were excluded. Body mass index, medication for high blood press- ure and housing tenure remained significant when they were added to the model together. Age and body mass index were positively associated with diastofic blood pressure. Women on medication for high blood pressure had higher diastolic blood pressure than those not on medication. Owner occupiers had lower diastolic blood pressure than those in rented acommodation. Both the baseline and final models had small but significant district level variance, representing about 2.6% and 2.4% respectively of the total variance. The models were also run without the variable "medication for high blood pressure" and the results (not shown) were similar. (b) Cholesterol Men. The variables years of education and hous- ing tenure were not significant, leaving a final model with age (which was not significant), social class and body mass index (Table 2). Cholesterol increased with body mass index and men in non manual social 'classes had higher levels of choles- terol. The district level variance was not significant in the baseline and final models by the criterion of inspection of the standard error. When the district level variance was tested using the likelihood ratio test against a model with the variance constrained to zero, it was found to b¢ significant in both the Table 2. Estimates from multilevel models of total serum cholesterol (in mmol/l) for men and women from the SHHS Men (n ffi 4647) Women (n - 4440) Baseline Final Baseline Final Fixed effects ¢.onsta~t age body mass index r.o~i~l ¢lase years of ~tion ~n~om eff~ lev~ I [~el 2 6.37 (0.025) 5.10 (0.13) 6.57 (0.03) 6.14 (0.15) 0.0055 (0.003) 0.005 (0.003) 0.088 (0.003) 0.083 0.003) 0.052 (0.005) 0.029 (0.004) oa4~ (0.035) -0.028 (0.008) 1.36 (0.028) 1.326 (0.028) 1.52 (0.032) 1.50 (0.032) 0.0074 (0.004) 0.0074 (0.004) 0.017 (0.007) 0,019 (0,008) Standard errors in parentheses. Level l~individual level. Level 2,--district level.

896 Carole Hart et al. Table 3. Estimates from multilevel models of alcohol consumption for men and women from the SHHS I I (i) Alcohol consumption modelled as units per wetk Men (n - 4890) Women (n - 4895) Baseline Final Baseline Final F~xed effec~ constant 17.2 (0.57) 18.83 (1.62) 4.59 (0.12) 3.12 (0.5) age . .-0.43 (0.05) -0.46 (0.05) --0.13 (0.02) -0A2 (0.02) social class -2.65 (0.67) 0.71 (0.2l year~ of education -0.18 (0.12) 0.14 (0.04) housing tenure -2.02 (0.65) 0.4 (0.2) Random effects level I 379.9 (7.7) 375.6 (7.61) 38.6 (0.78) 38.21 (0.77) level 2 5.31 (2.12) 4.05 0.74) 0.147 (0.097) 0.09 (0.08) (ii) Alcohol consumption modelled as non drinker versu~ drinker Fixed effectt constant -1.46 (0.06) -0.84 (0.23) -0.61 (0.06) -0.04 (0.19) age 0.04 (0.006) 0.03 (0.007) 0.05 (0.005) 0.04 (0.006) social clas~ -0.16 (0.09) -0.42 (0.07) years of education -0.05 (0.02) -0.05 (0,02) housing tenure -0.28 (0.08) -0.46 (0.07) Random effects level I I level 2 0.06 (0.028) 0.05 (0.025) 0.05 (0.02) 0.03 (0.015) 'Constrained by the method. ! i ! I Standard errors in parentheses. Level I--individual level. Level 2---district level. baseline and final models. The district level variance was, however, very small, representing only 0.5% of the total variance. Women The significant variables were similar to the model for men, with the exception of social class which was not significant and years of education which was significant. Cholesterol increased with age and body mass index and decreased with more years of education. There was a small but signifi- cant district level variance in both baseline and final models, representing over 1% of the total variance. (c) Alcohol consumption Units per weekmmen. There was a small but sig- nificant district level variance in the baseline model, representing 1.4% of the total variance (Table 3). Adding the significant variables social class, years of education and housing tenure decreased district level variance, but it remained significant and accounted for 1.1% of the total variance. With all the significant variables added together, the variable years of education became non significant. The model suggests that owner occupiers drink 2 units per week less than non owner occupiers, men in lower social classes drink more than men in higher social classes and older men drink less than younger men. Units per week--women. The district level var- iance was significant in the baseline model, using the likelihood ratio test but was not significant in the final model. The same variables were significant for women as for men and the model suggests that older women drink less than younger women. The coefficients of the other explanatory variables had opposite signs to the men, suggesting that women in higher social classes, women with more education and owner occupiers drink more. Women drink considerably less than men (average units per week were 17.3 for men and 4.6 for women for the 4890 men and 4895 women included in the alcohol model). Non-drinker--men. The baseline model had sig- nificant variance at the district level, representing 5.7% of the total variance. When other variables were added to the model, the district level variance was reduced but was still significant. The variable social class became non significant. The probability of being a non-drinker was higher for older men, lower for men with more education and lower for owner occupiers. Non-drinker--women. The results for women were similar to those for men, although social class remained significant in the model, and showed the probability of being a non-drinker was lower for women in non manual social classes. (d) Smoking Men. The baseline model had significant variance at the district level, representing 5.7% of the total variance (Table 4). After adding the significant explanatory variables social class, years of edu- cation and housing tenure, the district level variance was not significant (1% of the total variance). The probability of being a current smoker was higher for younger men, for those in manual social classes and for those renting and decreased with more years of education. Women. The results for women were similar to those for men. The district level variance in the baseline model was significant, but after adding explanatory variables, it became non significant (0.5% of the total variance).

.1 ! 1 i i i I I i I I I ! I i 1 People, places and CHD risk factors 897 Table 4. Estimates from multilevel models of current cigarette smoking for men and women from the SHHS, modelled as current cigarette smokers versu~ others Men (n = 4901) Women {n = 4913) Baseline Final Baseline Finat Fixed effec~ constant -0.46 (0.06) 0.63 (0.19) -0..~0 (0.06) 1.1 (0.19) age -0.01 (0.005) -0.03 (0.005) -0.01 (0.005) -0.03 (0.006) social class -0.38 (0.07) -0.31 (0.07) years of education -0.08 (0.01) -0.13 (0.02) housing tenure -0.76 (0.07) -0.77 (0.07) Random effects level I Ia la 1~ level 2 0.06 (0.02) 0.01 (0,01) 0.06 (0.02) 0.005 (0.008) "Constrained by the method. Standard errors in parentheses. Level l--individual level. Level 2---dlstrict level. Smoking was also defined as ever smokers of cigarettes, pipes or cigars versus others and the results (not shown) were similar. Analysis of residuals For some dependent variables which had siguifi- cant district level variance in their final models, the differences between each district and the whole data set were plotted for each baseline and final model. Figure 1 shows the district level residuals for models of diastolic blood pressure for men. In the baseline model, diastolic blood pressure varied from Renfrew (an urban industrial district) having nearly 3mmHg higher values than the overall average and Banff and Buchan (an agricultural dis- trict), Kirkcaldy (a manufacturing area), Stiriing (an agricultural district) and Nithsdale (a rural area) having 2mmHg lower diastolic blood pressure. Residuals for the final model were similar. Those districts with statistically significant residuals in the final model are marked with asterisks on the figures. For the female diastolic blood pressure baseline model (Fig. 2), the difference varied between 2.6 mmHg above the overall average in Roxburgh (a rural border district) and 3 mmHg below the average in both Kirkcaldy and Stiriing. Again, re- siduals for the final model were similar to the base- line model. In the models of cholesterol for females, six dis- tricts were close to the overall average (Fig. 3). Cholesterol varied from 0.3 mmol/l above average in goxburgh to 0.2mmol/1 below average in Glasgow South (an urban industrial area including city suburbs) for the baseline model and similar results were seen for the final model. Modets for male alcohol consumption (in units per week) showed variability from nearly 4 units per week more than average in Glasgow North (an Banff & Buchan* City of Aberdeen* City of Dundee Perth & Kinross Kirkcaldy* Dumfermline City of Edinburgh East Lothian* Roxburgh Stifling* Falkirk* Dumbarton Renfrew* Glasgow South* Eastwood Monklands Hamilton* Cunninghame Kyle & Carrink Nithsdale Glasgow North* -3 I'-"] baseline model ~ final model -I 0 1 2 3 Residual *Districts with significant residuals at 5% level in the final model Fig. I. Residuals from the multilevel models of diastolic blood pressure (mmHg) in men.

898 Carole Hart et al. Banff & 8uchan* City of Aberdeen" City of Dundee Perth & Kinross Kirkculdy~' Dunfermline City of Edinburgh East Lothian Roxburgh* Stifling* Falkirk Dumbarton Renfrew Glasgow South Eastwood Monklands Hamilton Cunninghame Kyle & Carrick Nithsdaln* baseline model final model Glasgow North* ~ ~ -4 -3 -2 -I 0 I 2 Residual *Districts with significant residuals at 5% level in the final model Fig. 2. Residuals from the multilevel models of diastolic blood pressure (nmaHg) in women. industrial inner city area) to 4 units per week below average in Perth and Kinross (a rural area) (Fig. 4). After including the other variables, the residuals from the final model became closer to the average for most districts, particularly in Eastwood (a sub- urb of Glasgow) and Glasgow North, but for the City of Edinburgh the residual in the final modct was larger than in the age only model. Further details of the characteristics of the dis- tricts are given elsewhere (TunstaiI-Pcdoc et ai., 1989). District level residuals were also plotted for the smoking models for males and females (Fig. 5 and Fig. 6). The residuals for the final models were much smaller than the residuals for the baseline models and none were significant. Inverness Banff & Buehan City of Aberdeen City of Dundee Perth & Kinross Kirkealdy Dunfermline City of Edinburgh East Lothian* Roxburgh* Stirling Falkirk Dumbanon Rcnfrew Glasgow South* ~,astwood Monkland$ Hamilton Cunninghame Kyle & Carriek Nithsdal¢ Glasgow North -0.3 -0.l 0 Residual baseline mode[ final model 0.1 0,2 0.3 *Districts with significant residuals at 5% level in the final model Fig. 3, Residuals from the multilevel models of total serum cholesterol (mmol/l) in women. ! I ! !

1 I 1 I I i I ! I I I 1 1 ! | People, places and CHD risk factors 899 Banff" & Buchan* City of" Aberdeen City of Dundee Perth & Kinross* Kirkcaldy Dunfcrm[in¢ City of Edinburgh* Eaat Lothian Roxburgh Stirling Falkirk Dumbarton Renfr©w" Glasgow South Eaatwood Monkland$ Hamilton Cunninghame Kyle & Carrick Nithsdnle Glnagow North* -5 baseline model final model -I Residual I 0 I 2 3 4 *Diatrictt with significant residuala at 5¢~ level in the final model Fig. 4. Residuals from the multilevel models of alcohol consumption (units/week) in men. DISCUSSION AND CONCLUSIONS The Scottish Heart Health Study was set up, in part, to further understanding of the geograp~cal variation in heart disease through a systematic cross-sectional study of the variation between areas in the known risk factors, such as diastolic blood pressure, cholesterol, smoking and alcohol con- sumption (Smith et al., 1987). This multilevel analy- sis has directly addressed this question by examining the effects of area on each risk factor separately, and for each sex separately, once a nurrt- ber of socioeconomic (housing tenure, social class and years of education) and other variables have been controlled for. Significant area effects after controlling for the socioeconomic and other explanatory variables, have been shown for three out of the four variables studied (namely diastolic blood pressure, cholesterol and alcohol consumption). Inclusion of the variable "medication for high blood pressure" in the blood pressure models was to ensure that the residual differences in blood Banff & Buchan City of Aberdeen City of Dundee Perth & Kinross[ Kirkcaldy Dunfermlion City of Edinburgh East Lothian Roxburgh S~irling Fnlkirk Dumbur~oo Renfrew Glasgow South Eastwood Monklands Hamilton Cunninghame Kyle & Carrick Nithsdnl¢ Glasgow North -0.4 -0.3 -0.2 -0.1 I ['---'l baseline model ~ final model 0 0.1 0.2 Residual 0.3 0.4 0.5 0.6 Fig. 5. Residuals from the multilevel models of current cigarette smoking in men.

Carole Hart et ~/. [nvnrness ~ Banff & Buchan City of Aberdeen [ City of Dundee Perth & ginross girkcaldy ~ baseline model Dunfermline ~ final model City of ~inburgh [ East Lothian Roxburgh Stifling Falkirk Dumbanon Rcafrew Glasgow South Eastwo~ ~ Monklands ] Hamilton Cunninghame Kyle & Carrick Nhhsdale Glasgow North ~ ~.5 -0.4 -0.3 ~.2 ~.I 0.I 0.2 0.3 0.4 0.5 Residual Fig. 6. R.iduals from ~e multiluel m~els of ¢~nt ci~tte ~oking in women. pressu~ we~ not du~ ~lely to differences betw~n of education and in owner o~upi~ ho~ng. distfic~ in the n~r of ~opl~ on m~i~fion for is in ~ntr~t to the pat~m for ~les. ~is differ- high blood p~u~. The~ may ~ differing levels ¢n~ ~tw~a th¢ sexes in the s~onomic pat- of ~ning for or awareness of high blood pres~ te~ng of al~hol consumption is ~n in other u~ at the pfima~ ~re l~el ~tw~n dis~cts, epid~olo~l studies including ~ WhitehaR II w~ch could lead to differing num~ on m~i- study (Ma~ot et al., t991). ca~ion. Excluding th¢ v~able "m~i~tion for ~ modelled together, as in Darien blo~ proud'" from the models did not aff~t the in.ration ~th ~x n~ds to ~ expli~y modell~, main r~ul~, although th¢ pe~en~ge of vafian~ and ~is ap~ not to have ~n done o~ufing at the dist~ct level w~ slightly ~gh~ above paper, ~th po~ible bias. in the estates of (2.6% for mal~ and 2.5% for femurs) which co~d ~¢ s~ of the area eff~t r~ulting. For chol.terol, suggest differen~ in num~ on mediation for diffe~nt soci~onomic variables ~gh blo~ pr~sur¢ ~twe¢n district. M~i~fion the ~nal model for women than ~at for men. has ~¢ ¢ff~t of r~ucing blood press~e, but tho~ In ~nt~st to the model of ~cohol consumption, oa m~i~tion a~ still ]i~¢ly to have hi~er blood ~¢ m~el of non~finking show~ v~ s~ar p~ure than nodal. ~is is ~n in the m~¢Is by ~ul~ for men and women. M~e~t¢ ~cohol con- the positive ~¢~ent of the "m~icafion for ~ s~ption is gene~lly associat~ ~th ~u~d cot- blood pre~ure" variable, ona~ h~ disease risk in epid~miolo~l studies, With the exception of smo~ng ~ ~n~1 ~or with ~th non~fink~ and heavy ~ns~e~ b~ng cxplanato~ vafiab1~ had relatively little cff~t on at inc~a~ ~sk (~ et al., 199~. ~Io~I the si~ of the r~siduals co~poading to inked- studi~ in the U~t~ Stat~ have generally shown r ual areas. Of ~e variables (ex~pt smokinD for lower corona~ h~rt di~a~ mo~li~ in smt~ which ~sidual variation is shown by ar~ ~e ~r- wi~ hi~er al~hol consumption (~hmidt and cen~ge change in ~is component of variation fob Popham, 1981; Laporte et al., 1980; Werth, 1980). lowing control for ~p]anato~ vafiabl~ vafi~ Therefore it is po~ible ~hat alcohol cons~ption from a ~duction of 24% (alcohol, men) to an coatfibut~ to popu]a~on differea~ in ~ron~ incr~ of 12% (cholesterol, women). ~¢ smo~ng hear ~s¢, althou~ in both ~ologi~l and indi- models showed ve~ large reductions i~ ~sidual vidu~-ba~d studi~s ~e possible contribution of vafian~ after control for individual cha~cte~ti~ ~nfoundiag facton has not b~n fully explored. (83% for males, 92% for femal~). As always with ¢pidemiolo~l studio, W~ pr.ent models for mal~s and females separ- cl~ions mch~ are limited by the variables in the ately. ~is allows the sep~at¢ ~timation of ar~ model ~d inde~ ofi~nally coll~t~. It is ~ible eff~s for mal~ and f~al~ as w¢II as allo~ng the ~at ~th the addition of fu~h¢r indi~du~ effects of ~¢ fix~ explanato~ variables to differ vafiabl~ in the model, the s~e and the ~gnifi~n~ b¢tween sexes. It is of not¢ that femal~ drink more of the ~a level ~ndom te~s would be r~u~d, when in non manual ~cial elates, ~ more yea~ though we should note that they do, in th~ ca~ of ! I I I I I

4. " 0.5 ng. This is differ- nic pat- n other II ~93), an todelled, in the nates of lesterol, uded in 1. .mption, similar aol con- :ed cor- studies, rs being ological shown 1 states dt and • 1980). tmption ,)ronary ad indi- tion of ~'edo lie con- ~ in the 3ossible educed, case of People, places and cholesterol for women, increase between the base- line and final models. The conclusions from such a study must be also be specific to the nature and size of the areas investigated, in this case local auth- orities. Some caution must also be exercised in making inferences to Scotland as a whole as the areas in the study were not selected at random. However a lot of thought was given to the selection of the areas, which reflect the complex nature of the known distribution of coronary heart disease mor- tality- in Scotland, with "a concentration of high mortality districts in the West of Scotland, but within that area two districts with very low mor- tality rates" (Smith et aL, 1987, p. 213). They also "reflect the diversity of Scottish life and correspond with recognisable population entities" (Tunstall- Pedoe et aL, 1989, p. 559). Moreover, the 22 dis- tricts were selected "to ensure good geographical spread and a range of urban/rural composition throughout Scotland... " (Smith et al., 1987 , p.213). In addition, the relatively large number of individuals sampled in each area means that indi- vidual residuals are quite often sufficiently large in relation to their standard error for them to be worthy of further investigation in their own right. The pattern of residuals found here is, as would be expected, in substantial agreement with local authorities noted as extreme in previous analyses (Tunstall-Pedoe et al., 1989). In particular Stirling, identified as having the lowest diastolic blood press- ure in men also has amongst the lowest residuals in both baseline and final models. Likewise, Renfrew has the highest diastolic blood pressure for men and also the highest residuals from the models. Analysing the significant level 2 residuals has enabled identification of particularly "healthy" or "'unhealthy" districts. Glasgow North has high re- siduals in the diastolic blood pressure models for men and women and the alcohol model (in units per week) for men. Conversely, Banff and Buchan has low residuals for these three models. It also has high residuals in the non-drinker models for men and women. Roxburgh has high residuals for two of the models for women (diastolic blood pressure and cholesterol) and Renfrew has high residuals for two of the models for men (diastolic blood pressure and alcohol in units per week). This study has contributed positively to the im- portance of area in addition to individual character- istics, by finding significant area effects for diastolic blood pressure, cholesterol and alcohol (males only). Physiological risk factors, such as diastolic blood pressure and cholesterol will reflect beha- vioural risk factors (such as animal fat consump- tion, salt consumption and exercise) which could be influenced by cultural effects of area of residence. Alcohol consumption amongst men could clearly have local cultural influences. Although the data were specific to Scotland, the results could be useful CHD risk factors 901 when attempting to reduce coronary heart disease in other places experiencing high rates. The residual effects of area are small and would only predict minor differences in coronary heart dis- ease risk. It should be remembered that in analyses like these though, the most marked residuals are likely to reflect measurement error (Efron and Morris, 1977). In these empirical Bayesian analyses, the residuals are shrunk towards the overall mean (Goldstein, 1995). A difference in 2 mmHg in dia- stolic blood pressure ~ould translate into a 9% difference in coronary heart disease risk on the basis of current evidence from prospective and intervention studies (MacMahon et aL, 1990; Collins et al., 1990). This blood pressure difference is greater than the majority of residuals seen and with methods taking into account the likely overes- timates of extreme differences is likely to be a near- maximal estimate of the actual area effects. Taking the difference between the greatest positive and greatest negative residuals of the statistically signifi- cant places as 4 mmHg, a 16% difference in coron- ary heart disease risk would be predicted. Taking 0.15mmoi/1 and 0.30 mmol/l cholesterol differences as equivalent area effects, a 3% and 7% difference in coronary heart disease risk would be predicted (Shepherd et al., 1995). Of all the baseline models• smoking showed the largest percentage of variance explained at area level. Figure 5 and Fig. 6 show clearly how the area effects in the baseline smoking models were mark- edly attenuated by adjustment for individual characteristics (social class, years of education and housing tenure). In conclusion, we have found relatively small area effects compared to individual effects for all variables examined, ie the majority of health effects are due to individual characteristics. This is in ac- cordance with a number of other health-related stu- dies using multilevel models (Humphreys and Carr- Hill, 1991; Duncan et aL, 1993; Ecob, 1996; Duncan et al., 1996). However, the significance of area effects in some of the models is evidence that places as well as people may play a role in coronary heart disease. Implications of this, which could be incorporated into health policy aimed at reducing coronary heart disease, suggest targeting character- istics of areas with high rates of coronary heart dis- ease, in addition to focusing on the indi~'iduals within these areas. Acknowledgements--This study was funded by the Scottish Office Home and Health Department and used data from the Scottish Heart Health Study Archive based at the University of Glasgow. The Scottish Heart Health Study was conducted by Hugh'TunstalI-Pedoe and col- leagues at the Cardiovascula/ Epidemiology Unit at Ninewells Hospital and Medical School, Dundee.

902 Carole Hart et aL REFERENCES Collins, 11., Peto. 11.. MacMahon, S., Hebert, P., Fiebach, N. H.. Eb~rl~in, K. A.. Godwin. J'., Qizilbash, N., Taylor. J. O. and Hennekens, C. H. 0990) Blood press- ure, stroke, and coronary heart disease: part 2, short term reductions in blood pressure: overview of random- ised drug trials in their epideminlogical context. Lancet 335, 827-838. Department of Health (1992) The Health of the Nation: A Strategy for Health in England. HMSO, London. Duncan, C., Jones, K. and Moon, G. (1993) Do places matter? A multi-level analysis of regional variations in health-related bchaviour in Britain. Social Science & Medicine 37, 725--733. Duncan, C., Jones. K. and Moon, G. (1996) Health-re- lated behavinur in context: a multilevel modelling approach. Social Science & Medicine 42, 817-830. Ecob, 11. (1996) A multilevel modelling approach to exam- ining the effects of area of residence on health and func- tioning. Journal of the Royal Statistical Society A 159, 61-75. Efron, B. and Morris, C. (1977) Stein's paradox in stato istics. Scientific American 236, 119-127. Goldstein, H. (1995) Multilevel Statistical Models. Edward Aruold, London. Humphreys, K. and Cart-Hill, R. (1991) Area variations in health outcomes: artefact or ecology. International Journal of Epidemiology 20, 251-258. Jones, K. and Moon, G. (1993) Medical geography: taking space seriously. Progresa in Human Geography 17, 51.5- 524. Laporte, R. E., Cresant~ .l.L. and Kuller, L. H. '(1980) The relationship of alcohol consumption to athero- sclerotic heart disease (review). Preventive Medicine 9, 22--40. Macintyre, S., Ma¢Iver, S. and Sooman, A. 0993) Area, class and health: should we b~ focusing on places or people?. Journal of Social Policy 22, 213-234. MacMahon, S., Peto, R., Cutler, J., Collins, R., Sorlle, P., Neaton, J., Abbott, R., Godwin, J., Dyer, A. and Stamler, J. 0990) Blood pressure, stroke, and coronary heart disease: part I, prolonged differences in blood pressure: prospective observational studies corrected for the regression dilution bias. lancet 335, 76~-774. Marmot, M. G., Davey Smith, G., Stansfeld, S., Patel, C., North, F., Head, J., White, I., Brunner, E. and Fccney, A. (1991) Inequalities in health twenty years on: the Whitehall II study of British civil servants. Lancet 337, 1387-1394. Rimm, E. B., Klatsky, A., Grobbce, D. and Smmpfer, M. J. (1996) Review of moderate alcohol consumption and reduced risk of coronary heart disease: is the effect due to beer, wine or spirits?. British Medical Journal 312, 731-736. Rodriguez, G. and Goldman, N. (1995) An assessment of estimation procedures for multilevel models with binary responses. Journal of the Royal Statistical Society A 1.~, 73-89. Schmidt, W. and Popham, 11. E. (198t) Alcohol consump- tion and isehemic heart disease: some evidenc~ from population studies. British Journal of Addiction 76, 407- 417. Scottish Office (1992) Scotland's Health: A Challenge to Us All: A Policy Statement. HMSO, Edinburgh. Shepherd, J., Cobb¢, S. M., Ford, I., Isles, C. G., Lorimer, A. R., Macfarlane, P. W., McKillop, J. H. and Packard, C. J. (1995) Prevention of coronary heat dis- ease wi~h prevastatin in men with hypereholesterolemia. New England Journal of Medicine 333, 1301-1307. Smith, W. C. S., Crombie, L K., Tavendale, R., I~ng, J. M., Kenicer, M. B. and Tunstall-Pedoe, H. (1987) The Scottish Heart Health Study: objectives and develop- ment of methods. Health Bulletin (Edinburgh) 45, 21 i- 217. Smith, W. C. S., Tunstall-Pedoe, H., Crombie, I. K. and Tavendale, R. 0989) Concomitants of excess coronary deaths--major risk factor and lifestyle findings from I0359 men and women in the Scottish Heart Health Study. Scottish Medical Journal 34, 550-555. Smith, W. C. S., Lee, A. J., Crombie, I. K. and Tunstall. Pedoc, H. (1990) Control of blood pressure in Scotland: the rule of halves. British Medical Journal 300, 981-983. TunstalI-Pcdo~, H., Smith, W. C. S., CYombie, I. K.. and Tavendale, R. (1989) Coronary risk factor and lifestyle variation across Scotland: results from the Scottish Heart Health Study, Scottish Medical Journal 34, 556- 560. Werth, J. (1980) A little wine for thy heart's sake (l~tter). Lancet ii~ 1141. Woodhouse, G. (ed.) (1995) A Guide to MLn for New Users. Multilevel Models Project. Institute of Education, University of London. I l l l l [ I [ 1 l l [ [ [ I