Multilevel Analysis Of The Geography of Health Inequalities: Using Better Geographies

Richard Mitchell, Simon Gleave, Mel Bartley

SSRU, City University, Northampton Square, London EC1V 0HB

Email: rm@ssru.city.ac.uk


Acknowledgements: This research is funded under the ESRC Health Variations Programme, Grant Number L128251012. The work is based on data provided with the support of the ESRC and JISC and uses boundary material which is copyright of the Crown and the ED-LINE Consortium. We gratefully acknowledge the ONS for access to the LS and Craig Duncan (University of Portsmouth) for access to the HALS Ward Identifiers.

1. Introduction

This paper is about factors which increase an individual's chances of being ill. Most studies which have investigated variation in health recognise that individual socio-demographic characteristics are important in determining a person's chances of having poor health. People in lower social classes (Pocock et al. 1987) the unemployed (Bartley, 1994) and women (Miles 1991) report and/or experience ill-health at higher rates, although recent work suggests these relationships are possibly not as straightforward as once thought (West 1997, Emslie et al. forthcoming). There is also increasing evidence that certain characteristics of the area in which an individual lives may have a significant role to play in mediating these individual level relationships. Shouls et al. (1997), Congdon et al. (1995), Haan et al. (1987), Gould & Jones (1996) and Ellaway & Mackintyre (1996) have all found area of residence and individual characteristics to be independently related to health (see Macintyre & Ellaway 1998 for a more comprehensive review of this work). The influences of area on an individual's health are usually referred to as 'area effects'. This paper will explore the relationship between individual and area level influences on health.

2. Why Do We Need To Include 'Different Geographies'?

The literature which describes the search for area effects is marked by contrasting results. For example, in contrast to those cited in the previous section, Sloggett and Joshi (1994,1998) found little or no impact of area characteristics on an individual's health once their own characteristics were accounted for. The research in general is also marred by poor or absent theoretical accounts of how and why the characteristics of an area might actually exert an influence on the health of its resident population (Macintyre et al. 1993). This has often resulted in a poor choice of variables with which to characterise an area and confusion over the appropriate spatial scale of analysis to adopt. There has for example, been some use of multivariate classifications or deprivation scores to try and capture something about the social context of an area (Shouls et al. 1995, Wiggins et al 1998). Whilst empirically this work has been successful in identifying a differential effect of 'area type' on individual health, it is rarely made clear what it is that these classifications are capturing which has an effect on health. Without an explicit consideration of how the characteristics of an area are expected to influence health, it is of little surprise that contrasting results have been obtained in the search for area effects.

This paper explores the relative influence of individual characteristics and area characteristics on health. Our question is, "does area of residence have an influence on an individual's health in addition to their individual socio-economic and demographic characteristics?". We also advance the theoretical approach to area effects with a focus on two different kinds of variable.

The first variable explores how people interact with their community. Whilst it is a relatively straight-forward task to identify aspects of an area's physical environment which might influence health, (through the biological effect of atmospheric pollutants for example), the effects of an area's social and economic milieu are much harder to identify. This is because those who are spatial neighbours are not necessarily social neighbours. Multivariate approaches to area 'profiling' attempt to describe a social or economic milieu through the identification of groups of similar types of people living together (Mitchell et al. 1998). These techniques cannot take account of the fact that people differ in their interactions with those who live close to them and therefore offer a problematic quantification of the social interactions in an area. In this paper, we have tried to explore how people's interactions with their community might influence their health through the inclusion of a variable which describes whether an individual feels part of their community, rather than one which describes the aggregate social profile of their spatial neighbourhood. Feeling part of the community can be conceptualised either as an area level variable (given that it may depend on how 'friendly' a place is) or an individual level variable (given that it may depend on how 'friendly' or outgoing the individual is). Here we assume that the variable has captured something about an individual's interaction with their social and economic habitat.

The second variable is a measure of the deindustrialisation which an area experienced between 1981 and 1991. Deindustrialisation is associated with economic decline and subsequent socio-demographic change which may in turn be associated with a fall in the chances of good health for those who remain resident in the area (Phillimore et al. 1994). This change might perhaps be caused by a reduction in the recreational opportunities available, the quality of health care available, the opportunities to eat properly and cheaply (Macintyre et al. 1993) or by increasing the stress in life (Bourdieu 1993). By choosing a variable with direct links to factors thought to decrease resident's chances of staying healthy or getting healthier, we hope to be able to make explicit links between residential location and health rather than make inferences based on multivariate descriptions of area.

The research strategy adopted was first to establish the nature and magnitude of relationships between the individual socio-demographic and economic variables and a health outcome measure. We then explored how attitudes to the community varied with the individual socio-demographic and economic variables and how it affected their relationships with the health outcome measure. Finally we introduced the area level variable and explored its relationship to the community interaction variable, and the health outcome measure. Both ordinary and multilevel regression analyses were employed to determine whether living in a deindustrialised area had a detrimental effect on health over and above individual level characteristics. The rest of the paper describes the models produced. In section 3, the data and techniques employed are introduced. Results from the individual analysis are reported in section 4. Relationships between individual characteristics and attitudes to the community are presented in section 5. The simple area analysis, using deindustrialisation and HALS is presented in section 6. Results from the multilevel analysis are reported in section 7 with discussion and conclusion following in sections 8 and 9.

3. Data And Methods

The evidence presented here has been drawn from two data sets, the ONS Longitudinal Study (LS) and the 1984/5 sweep of the Health and Lifestyle Survey (HALS). HALS sampled 9003 adults across Great Britain. It was designed to investigate a host of factors with an influence on health, and record the health of participants in both self-reported and clinically measured terms (Cox et al. 1987). In this research, the sample was reduced to 7818 respondents after eliminating those from Scotland, those with missing data and those with a social class other than the standard I - V. HALS allocated women a social class based on their husband if they were married and on their own occupation if not which meant that only a very small number of respondents were not allocated a social class. Scottish respondents were excluded because the LS only covers England and Wales and we wished to keep the HALS and LS analysis as comparable as possible. Age, sex, social class and time since employment were the basic individual level variables included in the HALS analysis. For this paper a measure of self-reported health has been used. It was derived from the number of 'physical' symptoms reported as affecting the respondent in the four weeks prior to interview (see table 1).

Table 1. Symptoms Included In The Measure

Headaches

Painful Joints

Constipation

Palpitations or breathlessness

Trouble With Eyes

Trouble With Ears

A Bad Back

Indigestion or other stomach trouble

Colds and Flu

Persistent Cough

Trouble With Feet

Faints or Dizziness

Kidney or Bladder Trouble

 

The measure was derived by summing the number of symptoms reported by the respondent. For some analyses, the variable was transformed into a dichotomous score, with a report of five symptoms or more placing the respondent in a 'high' number of symptoms category and a lower score being treated as a 'low' number of symptoms. 12.9 % of the sample reported 'high numbers of symptoms'.

Attitudes to the community were measured via response to a single question in HALS; "do you feel part of the community?". The question was posed following a number of others about the respondent's local area and we have therefore assumed that the intended definition of community is more heavily based on spatial neighbourhood than social neighbourhood. We recognise however that there is no way of knowing the definition of community which each individual respondent had in mind.

A definition of deindustrialisation was also required. This was defined as the fall in the proportion of a working population employed in manufacturing or energy related occupations, in a ward between 1981 and 1991(developing Champion and Townsend 1990). Thus, if a ward employed 60% of its workforce in manufacturing and energy related occupations in 1981, and 40% in 1991, it scored -20 on this scale. The deindustrialisation variable was calculated using the 1991 and 1981 census of population, with the 1991 data re-aggregated to 1981 ward level (Atkins et al. 1993). Since the HALS data were collected in 1984, the sample actually represents a population experiencing deindustrialisation rather than at the end of the process. It should also be noted that the data predate the job losses and community devastation which resulted from the 1984 miners strike. Thus the timing and the relative simplicity of our data may well distort the effects of deindustrialisation on health shown in the paper.

The relationship between deindustrialisation and the health of individuals was explored using two techniques. The first was to attach a measure of deindustrialisation to each of the individuals sampled in HALS, based on their ward of residence in 1984. Ward codes have been attached to the HALS respondents by Duncan et al. (1996), allowing a straightforward match. For tabulation and graphical purposes, deindustrialisation was transformed into a three class categorical variable. Class 1, 'Little Deindustrialisation or Growth' describes those wards which experienced only minor industrial employment losses, or which actually proportionally increased their industrial workforce (values greater than -10.6). Wards in class 2, ('Deindustrialised') experienced considerable falls in the proportion of industrial labour (values below -10.6, but above -20.0) and class 3, ('Extreme deindustrialisation') identifies those wards which suffered the greatest falls in proportion of industrial labour (values below -30.0). The deindustrialisation experienced by each individual was modelled as a contributing factor to the health outcome variable and tested for independence. The basic analysis was undertaken in SPSS for Windows, using the crosstabs procedure, linear regression and logistic regression. Some further tabular analysis was completed in EpiInfo.

Although they allow us to explore the effects of attitudes to community in relation to deindustrialisation, simple linear or logistic regression are not entirely appropriate for the hierarchical nature of the question we are asking. The second approach therefore was to use data from the LS and multilevel modelling. The LS is a 1% linked sample of individuals in England and Wales from 1971, 1981 and 1991 (Hattersley and Creeser, 1995). The dataset contains Census information and a limited health history from the National Health Service Central Register. We chose to employ LS data in the multilevel stage of the research for two reasons.

First of all, the LS offered a vastly improved sample size and avoids the spatial clustering present in the HALS sample; the analysis used a sample of 293,224 individuals aged 18 or over in 1981, who had full Census records in 1981 and 1991. This gave an average of 728 individuals in each local authority district. The greater sample size reduced concern about whether the HALS sample, which is clustered in urban areas, was truly representative of Great Britain as a whole. Although the LS has a much greater sample size than HALS, its non-clustered nature meant that analysis at ward level was not possible. Local Authority District level was used instead.

Secondly, the LS does not allocate women a social class based upon their husband's occupation. It therefore offered a chance to focus on the differences between social classes more effectively and on the effect of looking after the home and family. One disadvantage of the LS for this work is that the health outcome variable is self reported 'limiting long-term illness', rather than the measure of acute physical symptoms in HALS. The baseline relationships between individual socio-economic and demographic characteristics, and the health outcome variable are thus quite different to those described in HALS. A further disadvantage was that the LS contains no comparable variable describing an individual's attitude to their community.

Multilevel modelling allows both area and individual effects to be represented in a two level population hierarchy so an individual at level 1 is nested within area (in this case local authority district) at level 2. The hierarchical structure allows an understanding of how individual and area characteristics interact for every area, rather than attempting to model the relationship between area and individual characteristics with a single term as in logistic or linear regression. At the individual level, age, sex, social class and economic activity in 1981 were used to reflect an individual's circumstances within the labour market prior to the ten year period of deindustrialisation under investigation. Deindustrialisation was modelled as a continuous variable at the area level.

All the multilevel modelling was implemented using the software package MLn (Woodhouse 1996) which describes the appropriate statistical model for a binary outcome as a logistic multilevel regression model. In algebraic terms: -

yij = exp(fij + uj)/(1+exp(fij + uj)) + eij

where fij denotes the fixed part of the model, uj is the random part of the model at district level and eij is the random part of the model at individual level. The eij are assumed to be binomially distributed (Goldstein 1995). The fixed part of the model contains a linear function of both individual and area level explanatory variables. The random part of the model identifies two components of variance: that between area (level 2 variance) and that between individuals within area (level 1 variance). The inclusion of area level information into the model is equivalent to attempting to explain any between area differences. In this model, the only level 2 characteristic included is the deindustrialisation variable.

The modelling strategy was first to fit a simple variance components model to identify the components of variation. This is generally described as a null model as it does not contain any explanatory variables. The next step was to include an individual's age as age was expected a priori to be a predictor of limiting long term illness. The simple model with age and age squared is referred to as the base model. The next step was to include sex, social class, and employment status in the model and to check, by means of backward elimination, whether they make a statistically significant contribution to the model. This model is called the interim model. Finally, the level 2 deindustrialisation variable was added in an attempt to reduce area level variance. This is referred to as the final model.

This type of analysis is a breakthrough for data from the LS as previously, analysis of this type was only allowed on a limited set of variables read from a machine readable table. In the last year, ONS has allowed us the opportunity to download individual level data (under strict supervision) onto a standalone PC on which multilevel analysis can be carried out on larger and more interesting sets of data.

In the next section, results are presented from the analysis of HALS which establish the baseline relationships between individual socio-economic and demographic characteristics, and health. These are then followed by the results which describe how these relationships are changed in association with attitude to the community. The results are interpreted and discussed in the discussion section (see section 8).

4. Results: Baseline Relationships Between Health and Individual Level Variables (HALS)

Figure 1. Differences in The Reporting of High Number of Symptoms, by Age Group

Figure 1 shows a steady increase in the numbers of symptoms reported by men and women of all social classes, as they age.

Table 2. Differences in The Reporting Of High Numbers of Symptoms By Sex

Sex

% Sample Reporting
High No's Symptoms

N

Odds Ratio (With 95% CI)

Men

9.2

3384

1.0

Women

15.6

4434

1.82 (1.57 - 2.10)

Chi-sq. value = 69.7, p = .000

Table 2 shows a dramatic difference between the propensity of men and women to report high numbers of physical symptoms.

Figure 2. Differences in the Reporting of High Numbers of Symptoms By Social Class

Chi-sq. value = 48.645, p = .000

Figures 1-3 and table 2 show the relationships between health and age, sex and social class and the health outcome measure. These figures and the table describe the baseline relationships between individual characteristics and health which we expected to be altered by respondent's attitude to their community and by their area of residence. Figure 2 shows the class gradient which exists for the health of men and women of all ages. This relationship is likely to have been distorted by the allocation of their husband's social class to women. Since women are more likely to report high numbers of symptoms (as shown in table 2) and social class III non manual is in reality dominated by women, the graph should probably rise more steeply between classes II and III non manual. For the purposes of this paper however, we are interested in deviations from the baseline relationships rather than their precise nature.

Figure 3. Differences In The Odds of Reporting High Numbers of Symptoms By Length Of Time Since Employment

Figure 3 shows the relationship between ill-health and the length of time a respondent had been out of work. This variable includes the retired population as being non-employed, however as figure 3 shows, even after adjustment for age, the variable is a good predictor of ill-health. Further analysis showed that length of time 'not working' is much more powerful predictor of illness than unemployment or age alone.

The baseline relationships between age, sex, class, time since employment and health were confirmed as independent and significant by both linear regression and logistic regression. The next section describes the basic relationships between the individual characteristics and attitude to the community and then the relationship between health and attitude to the community.

5. Results: Who Feels Part Of Their Community And What Does It Mean For Their Health?

Figure 4. The Relationship Between Attitude to Community And Age

p = 0.05

22.9% of the HALS respondents included in the analysis did not feel part of their community. There were no significant differences in response between men and women, nor between social classes, but there was a significant relationship with age which is shown in figure 4.

Figure 4 suggests that in the earlier and mid adult years, when people are most likely to have recently moved and are perhaps most likely to socialise with those they work with, fewer people feel part of their community. As life becomes more stable, they move less and their children become increasingly independent, community involvement increases, especially around retirement.

Table 3. The Relationship Between Health Outcome and Attitude to Community

 

High No. Of Symptoms

Not High No. Symptoms

Relative Odds
(With 95% CI)

Feel Part of The Community

12.0 %

84.6 %

1.0

Do Not Feel Part Of The Community

15.4 %

88.0 %

1.34 (1.15 - 1.57)

Chi-sq. Value = 14.7, p value = .000

Table 3 shows that in the HALS sample, higher numbers of symptoms were reported by those who did not feel part of their community.

Table 4. Logistic Regression Showing The Relationship Between Health, the Individual Variables and Attitude to Community

 

Variable

B

S.E.

Sig

Exp B

Lower CI

Upper CI

       

These figures are odds ratios

Age

.0218

.0024

.0000

1.0221

1.0172

1.0269

Sex (men as baseline)

.5701

.0782

.0000

1.7685

1.5173

2.0614

Social Class

   

.0000

     

II (I as baseline)

.1492

.1971

.4491

1.1609

.7889

1.7081

III Non Man

.2634

.2052

.1993

1.3014

.8704

1.9458

III Manual

.5002

.1889

.0081

1.6491

1.1389

2.3880

IV

.6648

.1963

.0007

1.9442

1.3233

2.8564

V

.7153

.2265

.0016

2.0449

1.3118

3.1876

Time Since Employment

   

.0000

     

< 3 Months
(Employed as baseline)

-.0105

.2701

.9689

.9895

.5828

1.6800

4-6 Months

.2711

.2844

.3404

1.3115

.7511

2.2900

7-11 Months

.2844

.2555

.2657

1.3290

.8054

2.1929

1-2 Years

.2214

.1753

.2065

1.2478

.8850

1.7594

2-4 Years

.5160

.1259

.0000

1.6753

1.3090

2.1440

5-9 Years

.5686

.1169

.0000

1.7658

1.4041

2.2207

10-19 Years

.5831

.1239

.0000

1.7916

1.4053

2.2841

20 + Years

.4579

.1408

.0011

1.5807

1.1996

2.0830

Feel Part of The Community (Not part of community as baseline)

-.3173

.0803

.0001

.7281

.6221

.8522

Constant

-3.7702

.2289

.0000

     

Table 4 shows the relationships between age, sex, social class, the amount of time since employment, attitudes to the community and health. The results presented are from a logistic regression.

The table shows that feeling part of the community reduces the chances of reporting high numbers of symptoms. Notice too, that the detrimental effect of descending the social class scale only becomes significant at class III manual, and that for being non-employed starts 2 years after not working. If we use the number of symptoms as a continuous outcome variable and model it using linear regression, all the independent variables remain significant (p = .000).

This analysis demonstrates that feeling part of the community is significantly and independently related to the health outcome variable. The next section describes how it alters the baseline relationships between the socio-economic and demographic variables and health. Figures 5 and 6 and table 5 illustrate its effect with a comparison between the health gradients for those who do feel part of their community and those who do not.

Figure 5. The Relationship Between Age, Feeling Part of the Community and Reporting High Numbers of Symptoms

Figure 5 shows that amongst the middle - older age groups, not feeling part of the community is associated with an increased risk of reporting high numbers of symptoms.

Figure 6. The Relationship Between Social Class, Feeling Part of the Community and Reporting High Numbers of Symptoms

Figure 6 shows that for men and women of all ages, not feeling part of the community is associated with a higher risk of reporting symptoms, regardless of social class.

Table 5. Differences in The Reporting of High Numbers of Symptoms by Sex and Attitude to the Community

Feel Part of The Community?

Sex

Report High No.'s
Of Symptoms

N

Relative Odds
(95% CI)

No

Men

11.0 %

755

1.0

 

Women

18.7 %

1031

1.87 (1.34 - 2.18)

Yes

Men

8.6 %

2625

1.0

 

Women

14.6 %

3407

1.80 (1.52 - 2.15)

Table 5 shows that the relative differences between men's and women's health outcomes are preserved whether the respondents feel part of the community or not.

A useful way to conceptualise the health differences between those who feel part of the community and those who don't is this; for social classes I - III Manual, a respondent in a given class who feels part of the community will have the same propensity to report high numbers of symptoms as one who doesn't feel part of the community but who is in one social class higher. This difference is exacerbated for classes IV and V. Once over age forty four, the respondent who feels part of the community will have the same propensity to report high numbers of symptoms as one not feeling part of the community, but who is ten or fifteen years younger. Attitude to community does not therefore change the nature or direction of the baseline relationships between age and class and health, but it does dramatically change the health gradients associated with those variables.

In the next section of the paper, results are presented from HALS which illustrate how deindustrialisation is related to health and community attitudes.

6. Results: Who Experienced Deindustrialisation And What Effect Did It Have On Their Health? (HALS)

Using a basic mapping package, the spatial distribution of industrial decline described by the deindustrialisation variable can be revealed. Lines showing ward boundaries have been removed to give a better visualisation of the data.

Figure 7. The Spatial Distribution of Deindustrialisation 1981 - 1991 (Wards)

Figure 8. The Frequency of Values in the Deindustrialisation Index

Figure 8 shows the distribution and range of the deindustrialisation values mapped in figure 7. Notice that the average change in the proportion of industrial employment was -10.6 percentage points. The vast majority of wards in England and Wales experienced some loss of industrial employment.

The figures below show which HALS respondent groups experienced most deindustrialisation. The graphs show the presence or absence of each group of respondents in each of the three categories of deindustrialisation relative to the whole sample. Had there been no differences in group presence in districts with different levels of deindustrialisation, there would be no protrusion of the bars away from the X axis on the graph. For example, figure 9 shows that there were relatively more 25-34 year olds experiencing extreme deindustrialisation than we would expect had the HALS sample been evenly distributed amongst the three deindustrialisation groups. It must be noted that these graphs are not a comment on how deindustrialisation affected Britain's population, because the HALS sample was spatially clustered. The graphs are an indicator of which groups in this sample have experienced deindustrialisation.

Figure 9. Deindustrialisation Experiences of the HALS Sample, by Age Group

p = .040

Figure 10. Deindustrialisation Experiences of the HALS Sample, by Social Class

p= .000

Figure 10 suggests that amongst the HALS sample those in social class III manual have experienced greatest amounts of deindustrialisation. There were no significant differences in the likelihood of male and female respondents to have experienced different degrees of deindustrialisation nor for those out of employment for different lengths of time.

No significant differences were found between the likelihood of respondents experiencing different degrees of deindustrialisation to feel part of their community.

The next section describes the relationship between deindustrialisation, individual characteristics and the health outcome variable.

Figure 11. Differential Reporting Of High Numbers of Symptoms by Age-Deindustrialisation Group

p = .000

Figure 11 shows the relationship between deindustrialisation and health. It shows the proportions of respondents reporting high numbers of symptoms in each deindustrialisation class, relative to the whole population in each age group. Again, if there were no health differences between the deindustrialisation classes, the bars on the graph would not protrude away from 0 on the X axis. Figure 11 shows clearly the effect of deindustrialisation on the health gradients associated with getting older. In the middle and older-aged groups, deindustrialisation (especially in its most extreme forms) is associated with a larger increase in the likelihood of reporting high numbers of symptoms than that which we would expect.

Figure 12. Differential Reporting Of High Numbers of Symptoms by Class-Deindustrialisation Group

p = .000

The relationship between health and class in areas with different degrees of deindustrialisation is shown in figure 12. Class I appears to offer protection against the ill-effects of deindustrialisation. All the other classes show a marked increase in the propensity to report high numbers of symptoms in areas which have experienced extreme deindustrialisation. There is a notable peak in the effects of extreme deindustrialisation for class III non manual. This was a surprising result, possibly connected to the gender composition of the class, although given that HALS defines a women's social class by that of her husband if she is married, this is rather difficult to explore further.

Table 6. Logistic Regression Showing the Relationship Between Health, the Individual Level Variables and Deindustrialisation

Variable

B

S.E.

Sig

Exp B

Lower CI

Upper CI

Age

.0224

.0024

.0000

1.0227

1.0178

1.0275

Sex (Men as Baseline)

.5711

.0782

.0000

1.7702

1.5185

2.0635

Social Class

   

.0001

     

II (I as baseline)

.1492

.1972

.4495

1.1609

.7887

1.7087

III Non Man

.2467

.2054

.2299

1.2798

.8556

1.9143

III Man

.4419

.1896

.0198

1.5557

1.0728

2.2561

IV

.6111

.1970

.0019

1.8424

1.2524

2.7104

V

.6597

.2272

.0037

1.9342

1.2390

3.0195

Time Since Employment

   

.0000

     

< 3 Months
(Employed as baseline)

-.0141

.2702

.9585

.9860

.5806

1.6745

4-6 Months

.2654

.2844

.3506

1.3040

.7468

2.2768

7-11 Months

.2748

.2556

.2822

1.3163

.7976

2.1723

1-2 Years

.2259

.1754

.1978

1.2535

.8888

1.7678

2-4 Years

.5019

.1261

.0001

1.6518

1.2902

2.1148

5-9 Years

.5519

.1171

.0000

1.7365

1.3803

2.1846

10-19 Years

.5654

.1243

.0000

1.7602

1.3797

2.2457

20 + Years

.4562

.1410

.0012

1.5781

1.1971

2.0802

Feel Part of the Community (Not Part of The Community as baseline)

-.3139

.0804

.0001

.7306

.6241

.8553

Deindustrialisation

-.0194

.0048

.0000

.9808

.9716

.9900

Constant

-3.9770

.2352

.0000

     

Table 6 shows that deindustrialisation is independently and significantly related to the chances of reporting high numbers of physical symptoms. Its B value is negative because a greater degree of deindustrialisation is represented by a larger negative score. Hence, as the deindustrialisation score rises towards zero, the likelihood of reporting high numbers of symptoms reduces.

Linear regression confirms that this area level variable is also independently and significantly related to the total number of symptoms reported (p = .000).

In the next section, the results from multilevel analysis of the LS are presented.

7. Results: Multilevel Analysis Of Health And Deindustrialisation (LS)

Figure 13 shows the spatial distribution of deindustrialisation at the local authority district level. Note the effect of aggregation is to reduce the apparent severity of the loss of industrial employment. The average figure for districts is just -4.75, compared to -10.6 for wards.

Figure 13. Spatial Distribution of Deindustrialisation (Local Authority Districts)

Average = -4.75

Interpretation of the modelling is a two stage process. First of all, comparison of the log likelihoods before and after the inclusion of individual characteristics illustrates how far differences between areas are explained by differences between the individuals they contain. Secondly, the analysis of the residuals reveals the extent to which an area has an excess of ill-health (or good health) once individual characteristics and the deindustrialisation variable have been added.

Table 7. Beta Values For Baseline, Interim and Final models

Base Model

Interim Model

Final Model

FIXED EFFECTS

     

(Level 1)

     

Constant

-1.836 (0.016)

-2.041 (0.017)

-2.144 (0.022)

Age

0.066 (0.001)

0.066 (0.001)

0.0663 (0.001)

Age2

0.00012 (0.00002)

0.000036 (0.000023)

0.000037 (0.000023)

Female (relative to male)

 

-0.180 (0.014)

-0.180 (0.014)

Unemployed 1981(relative to employed)

 

0.652 (0.024)

0.652 (0.024)

Sick 1981 (relative to economically active)

 

2.952 (0.038)

2.954 (0.038)

Retired 1981

 

0.116 (0.021)

0.116 (0.021)

Look after home 1981

 

-0.049 (0.027)

-0.049 (0.027)

Other Econ Act 1981

 

0.839 (0.142)

0.842 (0.142)

Skilled Non Manual 1981

 

-0.027 (0.019)

-0.027 (0.019)

Partly Skilled/Unskilled '81

 

0.334 (0.015)

0.333 (0.015)

No Social Class 1981

 

0.467 (0.025)

0.467 (0.025)

       

(Level 2)

     

Deindustrialisation

   

-0.021 (0.003)

       

RANDOM EFFECTS

     

Level 1*

1

1

1

Level 2 (county district)

0.083 (0.007)

0.063 (0.005)

0.054 (0.005)

       

Log Likelihood

172,143

144,981

144,949

 

* Constrained by logistic multilevel model to be 1.
Std Errors In Brackets

Table 7 shows only those variables which remained significant at each stage of the modelling process. The values in the table are Beta coefficients and may be read in the same manner as those for logistic regression. From the baseline model, a positive relationship between an individual's age and the probability of reporting a limiting long-term illness is clearly confirmed. The quadratic term in age is also useful as health determinants tend to have stronger effects at older ages. Extrabinomial variation was tested for and found not to be present (Woodhouse 1996). The base model reveals that between-area differences are present (level 2 variance is 0.054) and the model shows that by including individual level characteristics, the area level variance is reduced by approximately a quarter. Differences between the aggregate health of districts, thus decrease once we take account of the characteristics of the individuals who make up the district populations.

Attempts to explain the remaining between-area differences, by taking account of the deindustrialisation have produced a reduction in level 2 variance by 14.28%. The fixed part of the model reveals that being sick, unemployed, retired or being in 'other economic activity' in 1981, in combination with being in a low class or having no social class at all increases an individual's chances of reporting a limiting long term illness. This risk is reduced if the individual is female and/or looking after the home.

Of greatest interest are those places which, from the analysis of residuals in the model, appear significantly more or less healthy, once individual and area characteristics are accounted for. As the modelling progressed the number of areas with a significant residual reduced. In the base model, where only differences in age had been accounted for, 73 significantly healthy and 88 significantly unhealthy districts remained. Once individual's economic status and social class were included in the model, this numbers fell to 52 healthy and 79 unhealthy districts. With the addition of deindustrialisation 51 significantly healthy and 69 significantly unhealthy areas remain unexplained.

Figure 14. Level 1 and Level 2 Models Of Limiting Long-Term Illness

These changes are shown in figure 14. The left-hand map shows districts whose population remains significantly more sick or more healthy, after we account for differences between the socio-economic and demographic characteristics of their populations. The right-hand map shows districts whose population remains significantly more sick or more healthy, after we account for differences between the socio-economic and demographic characteristics of their populations and the deindustrialisation which took place there.

Figure 15. The Impact Of Deindustrialisation On Health

To make these results easier to interpret, figure 15 shows those places in which there is apparently better health than we would expect, given the amount of deindustrialisation experienced (after accounting for the socio-economic and demographic composition of the population) and those places in which health is worse. It is a map of the effect of deindustrialisation on the health of the resident population. That the populations of some areas appear to experience better health once deindustrialisation is accounted for is slightly counter-intuitive.

In the next section of the paper, the results presented here are drawn together and discussed.

8. Discussion

As well as the expected relationships of ill-health to social class, sex and non-employment (figures 1,2 & 3 and table 2), this analysis has shown the importance of integration into one's community. Feeling part of the community was itself related to age (figure 4), but did not differ between men and women, or between social classes. The idea that quality of relationship with one's residential area has an impact on one's life and health is not new (see for example, Kawachi et al. 1997 or Tijhuis et al. 1995) but we had no a priori expectations about the relationship between an individual's sex and their attitude to the community. We were unsure if men would feel more part of their community because of the nature of their work or if women would feel more part of their community through the nature of their friendships. For social class however, the results seem to have exploded the myth (and our own expectations) that there is a greater degree of community coherence amongst working class groups or perhaps that the middle and professional classes lead more detached and private lives. It must be remembered that despite the position of the community question in the interview schedule, the definitions which were used in responses may have varied between a spatial, social or business based community. It may be that most people feel part of a community, but that the nature of community differs from group to group.

The relationship between the community variable and the health outcome variable is however, very clear. People who do not feel part of their community are more likely to report high numbers of physical symptoms. With cross-sectional data, the direction of the relationship (table 3) was difficult to determine. Do people with more illness not feel part of their community because their symptoms prevent them from taking an active role in it? Or are people who do not take an active part in their community somehow denied the health benefits of that interaction? It was not possible to provide an answer from these data and there is therefore a need for caution in our treatment of the result. We stress that there is an association between ill-health and not feeling part of a community, although logistic regression confirms that age, sex, class, time since employment and 'feeling part of the community' are all independently related to higher numbers of symptoms (table 4).

Community interaction had a different relationship to health at different ages (figure 5) and in different classes (figure 6), so that it acted to sharpen both the age and class gradients in health (see Stansfeld et al. 1998 for a further exploration of these concepts). Attitude to, and as we have assumed interaction with, the community, does therefore appear to play a significant part in individual's experiences of ill-health.

The fact that feeling part of the community remained independently related to health when deindustrialisation was included in the model marks a key aspect of this research. We have demonstrated that the relationship between deindustrialisation (an area characteristic) and health (an individual characteristic) is modified by the individual's attitude to their community (table 6). This represents an improvement over approaches which implicitly assume that individuals are part of a local social and economic milieu because their residential postcode locates them within an areal unit with a particular set of aggregate socio-economic characteristics (Beaumont, 1991). It is not sufficient to know what kinds of people an individual lives nearby, how those people interact with each other is also important.

This point highlights the fact that quantitative spatial analysts must not only keep pace with the technological developments around them, but also with theoretical developments which might enhance the variables and the concepts which are fed into the new technologies. This paper has shown that for those interested in the relationship between area and health, a consideration of how individual's use the place and space around them is as important as their residential location (note the B values for deindustrialisation and feeling part of the community in table 6 for their relative contributions to the model).

There has been much discussion in the media of the effects of deindustrialisation on communities and their health, aspirations and beliefs. The industrial decline of Britain in the 1970's and 1980's has recently been recalled by the film industry. Films such as "The Full Monty" and "Brassed Off" dealt directly with the social and health effects of the decline of heavy industry and mining, on the community and the individuals within it. The statistics which describe deindustrialisation are staggering. During the period 1981 - 1991, in the coalfields alone, nearly 160,000 jobs were lost (Beatty and Fothergill 1996). Champion and Townsend (1990) report a net loss of 2.8 million manufacturing jobs in Britain between 1971 and 1989, remarking that deindustrialised communities were not only faced with redundancies, but vast falls in recruitment rates. Families traditionally supported by the industries simultaneously lost the incomes of father and the employment prospects for son. Since, for economic and production reasons, the manufacturing and mining industries were spatially concentrated and surrounded by communities almost entirely dependent on them for work and economic life-blood, the effects of their loss were socially and spatially concentrated too. Analysis of the HALS sample (figures 9 and 10) suggest that it has captured those experiencing different degrees of deindustrialisation without great bias. People in skilled manual labour, at the early stages of their working lives appear to have experienced greatest industrial decline.

Before we consider the relationship between deindustrialisation and health, that between an individual's attitude to their community and their experience of deindustrialisation provides an intriguing finding. We wondered whether the experience of deindustrialisation might drive a community together, making the respondents more likely to feel part of a community united in 'hard times'. This would have resulted in more respondents saying that they felt part of the community in those wards which underwent greatest industrial decline. Or, perhaps the experience of industrial decline, job loss and possible industrial conflict would reduce the community coherence in areas which had experienced the harshest changes. Neither of these scenarios appeared to be correct, with no significant differences in attitude to the community between groups which experienced different degrees of deindustrialisation. It may be that the longer-term effects of industrial decline had yet to take an effect on the communities sampled for HALS.

With regard to health, deindustrialisation compounded an already considerable burden on the populations of heavy industrial workers. In addition to the physical effects of working in heavy industry, the populations had then to deal with unemployment and social and economic decline. It is this decline which has been pinpointed as a possible source of influence over health outcomes for the entire population of a deindustrialised area. It must also be remembered that the people experiencing industrial decline were often not very healthy before the process began.

The single-level analysis has shown that when people reach ages at which they naturally start to experience more ill-health, the industrial decline in their area appears to be exacerbating their own health decline (figure 11). If this were an effect of having worked in or close to heavy industry, we would expect the influences of deindustrialised and extremely deindustrialised areas to be similar. The fact that the greatest increases in ill-health accompany the greatest industrial decline suggests that this is a function of changes in the social milieu and the resources available to maintain health in those areas. There is also an interesting effect on the youngest age group in that those exposed to extreme deindustrialisation appear to have relatively higher chances of reporting high numbers of symptoms. A declining area which offers few job prospects and is perhaps struggling to support it's entire community might well have little opportunity to invest in the health of its young people. It may also be that some form of migratory selection is illustrated by these results with those young and healthy enough to do so, moving away from areas experiencing extreme deindustrialisation.

The single-level analyses we have presented suggest that even after we account for the effects of different personal circumstances, those living in an area which has experienced industrial decline are more likely to report higher numbers of symptoms (table 6). Being in a lower social class (III manual or below), being a woman and having been out of work for two years or more are also significantly associated with a greater chance of ill-health.

The differences between limiting long-term illness, used in the multilevel analysis and the measure of acute physical symptoms used in the analysis of HALS, are clearly shown by table 7. Here we find that in contrast to HALS, the LS suggests being a woman decreases one's chances of reporting a limiting long-term illness, and that the detrimental effect of being in a lower social class is only found in the lowest classes. These results convey the considerable differences between the two health outcome measures. There will also be some difference in the results obtained when local authority districts, rather than wards, form the spatial units for analysis. Given our implicit hypothesis that individuals draw influences over their health from their community, ward is probably a more appropriate scale to explore the effects of industrial decline. Methodological constraint has prevented us from applying multilevel techniques at this scale. We must therefore recognise that we may be masking much area effects by working with larger, less appropriate, spatial units.

The left hand map in figure 14 shows a clear Northwest / Southeast divide in health once individual socio-economic and demographic characteristics are accounted for. It is these differences between districts which medical geographers and sociologists wish to explain by identifying area characteristics with an impact on the health of residential populations. Table 7 shows that the inclusion of deindustrialisation explains some of the differences in health outcomes between populations living in different local authority districts. Figure 15 is the easiest reference point for understanding this argument. The areas highlighted in either red or blue are those in which deindustrialisation makes a significant contribution to explaining the health of the residential population, once we have accounted for their socio-economic and demographic characteristics. Given their individual characteristics, people in the blue districts are more likely to report an illness than they might have otherwise been without exposure to industrial decline.

The blue areas, where deindustrialisation has increased the risk of ill-health, did not necessarily experience the highest levels of industrial decline as we have defined it. It is however, important to remember the limitations of the technique and data in question. Deindustrialisation was not confined to the decade for which we have data, nor to those industries via which we have defined it. The finding that the places where health was apparently most affected by the decline are not those in which it was most severe is interesting and we suspect that there are two factors which explain it. First of all, many of the districts which have suffered greatest decline contained a very unhealthy population at the start of the 1980's. Extreme deindustrialisation in South Wales for example, does not appear to have explained the concentrations of ill-health found there (see figure 14). It seems likely that the population were already suffering the effects of heavy industrial and mining jobs and that deindustrialisation has had little additional effect. Secondly, the districts with most extreme industrial decline may have taken remedial action to offset its worst effects. Townsend and Peck (1985) emphasise the importance of understanding the interaction between local context and national or international events and policies when reviewing the effect of industrial decline on a particular area.

The areas in red in figure 15 are those with populations that are less healthy than we would expect having taken account of deindustrialisation. There are two factors which explain this result. Firstly, they are all places which experienced either small industrial growth between 1981 and 1991 or far less decline than the national average (compare the red districts on figure 15 with figure 13). Whilst severe industrial decline is associated with worsening health in an area, a small degree of industrial growth is not associated with improvements in the population's health. These places have achieved statistical significance because, relative to the rest of the sample, their populations have experienced industrial growth without a corresponding improvement in health outcomes. Secondly, not all forms of industrial decline are included in our index. Plymouth (see label on figure 15) for example, experienced considerable industrial decline in the period 1981-1991 through the closure of naval and merchant ship yards and other military cut-backs. However, since we are measuring deindustrialisation on the basis of manufacturing and energy related occupations alone, Plymouth's decline is not appropriately modelled. This leads to its identification as an area which, given the amount of decline it registers on our index, is much less healthy than we would expect. Plymouth's health problems have stemmed from a source other than the loss of manufacturing and energy related industry.

This example does not however detract from the fact that, given our definition of industrial decline, we have shown that the deindustrialisation experienced in an area has had an independent, significant and detrimental affect on health.

9. Conclusions

This paper has shown that attitude to (and perhaps therefore, interaction with) the community has an effect on health which is independent of other individual characteristics and the degree of deindustrialisation in an area. This suggests that research which tries to classify areas, and their resident populations, in order to explore the impact of the social and economic milieu needs to take account of how people interact with their communities. The geography of area effects in ill-health is not defined purely by its spatial characteristics, but also its social characteristics.

This paper has shown that living in an area in which deindustrialisation has taken place will significantly increase an individual's chances of reporting ill-health, measured both in acute and chronic terms. It has also shown the value of multilevel modelling in identifying the extent of that effect, although it should be recognised that the single-level analysis was able to correctly identify the same effect of industrial decline on health. The paper has made a useful contribution to the field by illustrating that there is a difference in the way that people interact with their communities and that this is associated with a difference in health outcomes. In effect the inclusion of a variable which describes attitudes to the community is an attempt to model the difference between the area in which an individual resides and the place in which they live (Graham 1998, Macintyre et al. 1993). Place, rather than space is the location in which an individual is exposed to different influences on their health and health-related behaviour and there is an urgent need for a better range of data sources which describe places rather than spaces. Supply of these kinds of data will require a new approach to the choice of spatial scale for analysis and the recognition that different groups in society interact socially and spatially with each other in different ways.

The paper has illustrated how new techniques and new concepts can be combined to provide an enhanced analysis of a complex relationship. It would appear that health is a function both of individual characteristics and area of residence.

References

Atkins, D., Charlton, M., Dorling, D. & Wymer C. (1993) Connecting the 1981 and 1991 Censuses, NERRL Research Report 93/9, CURDS: Newcastle University

Bartley, M. (1994) Unemployment and Ill Health - Understanding The Relationship. Journal of Epidemiology and Community Health, Vol. 48, No. 4 pp. 333-337

Beatty, C., Fothergill, S. (1996) Labour Market Adjustment in Areas of Chronic Industrial Decline. Regional Studies Vol. 30, No. 7, pp 627-640

Beaumont, J. (1991) Concepts and Techniques in Modern Geography: 53. An Introduction to Market Analysis. Environmental Publications: Norwich

Bourdieu P. (1993) La Misere Du Monde, Edition Du Seuil: Paris.

Champion, A., Townsend, A (1990) Contemporary Britain : a geographical perspective London:Edward Arnold

Congdon, P., Shouls, S., Curtis, S. (1997) A multilevel perspective on small-area health and mortality: a case study of England and Wales. International Journal Of Population Geography, Vol. 3., pp 243-263

Cox, B.D. (1987) The Health and Lifestyle Survey, HPRT: London

Duncan, C., Jones, K., Moon, G. (1996) Health-Related Behaviour In Context: A Multilevel Modelling Approach. Social Science Medicine, Vol. 42, No. 6, pp 817-830

Ellaway, A., Macintyre, S. (1996) Does Where You Live Predict Health Related Behaviours? A Case Study In Glasgow. Health Bulletin, Vol. 54, No. 6, pp 443-446

Emslie, C., Hunt, J., Macintyre, S. (forthcoming) Problematising gender, work and health: the relationship between gender, occupational grade, working conditions and minor morbidity in full-time bank employees. Social Science and Medicine

Gleave, S., Bartley, M., Wiggins, R. (1998) Limiting Long-term Illness: a question of where you live or who you are? LS Working Paper No. 77, SSRU: City University

Goldstein, H. (1995) Multilevel Statistical Models (2nd Edition), London, Edward Arnold; New York, Halstead Press

Gould, M., Jones, K. (1996) Analysing perceived limiting long-term illness using UK census microdata. Social Science and Medicine, Vol. 42, No. 6, pp 857-869

Graham, E. (1998) Placing Health. Paper presented to the ESRC Health Variations Programme, May 1998.

Hann, M., Kaplan, G., Camacho, T. (1987) Poverty and Health - Prospective Evidence From The Alameda County Study. American Journal of Epidemiology, Vol 126, No. 6, pp 989-998

Hattersley, L., Creeser, R. (1995) Longitudinal Study 1971 - 1991: History, organisation and quality of data. HMSO

Kawachi, I., Kennedy, BP., Lochner, K., ProthrowStith, D. (1997) Social capital, income inequality, and mortality. American Journal Of Public Health, Vol.87, No.9, pp.1491-1498

Macintyre, S., Maciver, S., Sooman, A. (1993) Area, class and health: should we be focusing on places or people?. Journal of Social Policy, Vol. 22, No. 2, pp 213-234

Macintyre, S., Ellaway, A (1998) Ecological Approaches: Rediscovering The Role Of the Physical and Social Environment. In Berkman, L., Kawachi, I. (eds) Social Epidemiology, Oxford University Press: Oxford

Miles, A (1991) Women, Health and Illness. Open University Press: Milton Keynes

Mitchell, R., Martin, D., Foody, G.M. (forthcoming) "Unmixing Aggregate Data: Estimating the Social Composition Of Enumeration Districts", Environment and Planning A.

Pocock, S., Shaper, A., Cook, D. (1987) Social Class Differences in Ischaemic Heart Disease in British Men. Lancet, Vol ii, pp 197-201

Phillimore, P., Beattie, A., Townsend, P. (1994) Widening Inequality Of Health In

Northern England, 1981-91, British Medical Journal, 1994, Vol.308, No.6937, pp.1125-1128

Shouls, S., Congdon, P., and Curtis, S. (1995) Geographic Variation In Health: The Development Of A Relevant Area Typology for SAR Districts

Sloggett, A., Joshi, H. (1994) Higher Mortality in deprived areas: community or personal disadvantage? British Medical Journal 1994, Vol 309, pp 1470-1474

Sloggett, A., Joshi, H. (1998) deprivation indicators as predictors of life events 1981-1992 based on the UK ONS longitudinal study. Journal of Epidemiology and Community Health, Vol. 52, No. 4, pp 228-233

Stanfeld, S., Bosma, H., Hemingway, H., Marmot, M. (1998) Psychosocial work characteristics and social support as predictors of SF-36 health functioning: The Whitehall II study. Psychosomatic Medicine, Vol. 60, No. 3 pp. 247-255.

Tijhuis, M., Flap, H., Foets, M., Groenewegen, P. (1995) Social Support And Stressful Events In 2 Dimensions - Life Events And Illness As An Event. Social Science & Medicine, Vol.40, No.11, Pp.1513-1526

Townsend, A., Peck, F. (1985) An Approach To The Analysis Of Redundancies In The UK. Chp 3 In Massey, D., Meegan, R. (eds) Politics and method : contrasting studies in industrial geography London:Methuen

West, P. (1997) Health Ineqaulities In The Early Years: Is there Equalisation in Youth? Social Science and Medicine, Vol 44. No 6. pp 833-858

Woodhouse, G. (1996) Multilevel Modelling Applications: A guide for users of Mln. London, Multilevel Models Project, Institute of Education