Volume 23
Number 3
2002

[Table of Contents]

Small area comparisons of health: Applications for policy makers and challenges for researchers

Paul J Veugelers and Shane Hornibrook

Abstract

It is a challenge to researchers to present their results in a way that serves the needs of health policy makers. Small area maps of life expectancy provide an insightful presentation. In this study, we pursued small area comparisons on a scale that is smaller than is currently available on a province-wide basis. We visualized Nova Scotia's provincial variation in health and identified the Cape Breton Regional Municipality and Halifax's disadvantaged “North End” neighbourhood as areas with major health concerns. The observed health differences are only partially explained by socioeconomic factors such as income and unemployment. The study also demonstrated the feasibility of small area comparisons at the level of census consolidated subdivisions and neighbourhoods. There are various methodological challenges for researchers, however: allocation procedures such as the postal-code-conversion-file may introduce substantial error; the application of appropriate spatial smoothing procedures is crucial to the interpretation of regional variation in health; and the migration of frail individuals to nursing homes affects the geographic variation in health.

Key Words: Bayesian methods; chronic diseases; health inequalities; health policy; life expectancy; mapping of disease, multilevel analysis; socioeconomic factors; spatial statistics

Introduction

Health inequality is a major public health issue.¹ Health inequalities occur in all kinds of geographically and socio-economically determined subgroups.^2,3 Their identification is crucial to health policy makers for planning and prioritizing prevention and intervention activities. The Government of Nova Scotia addressed health inequalities in its policy plan, stating: “fairness to regions and to people” and “where you live must no longer determine what you can be”.⁴ The principle of fairness extends to equal rights to good health among regions. To make this fairness a reality, provincial health inequalities should be monitored and reported in a way that is meaningful to policy makers.

Health policy makers are not necessarily trained as epidemiologists or statisticians

and may not have a thorough understanding of the results reported by researchers. Researchers are challenged to present their results in ways that serve the needs of health policy makers.^5,6 Various researchers are using geographic maps in which disease and mortality differences are visualized through colour patterns as a means of presenting their research results.^7–9 It would further ease interpretations if these colour patterns translate into quantities that are easy to comprehend: for example, if they translate into life expectancy rather than into standardized mortality rates or ratios.

Traditionally, geographic comparisons of health have used countries, states and provinces as their geographic units.¹⁰ These comparisons may fail to disclose the health concerns of smaller areas and are therefore not meaningful to health policy makers who govern such small areas.^11,12 It is for this reason that in past decades increasing numbers of studies have been investigating the health inequalities of smaller geographic units.^8,10–12 These small area studies have benefited from improvements in computers, geographic information systems, and statistical methods,8,10,13,14 but remain hampered due to concerns about the accuracy of the population information of small areas.^8,15 In addition, ecological bias resulting from selective migration is a larger concern in small area studies since intra-regional migration affects small area comparisons to a larger extent than inter-regional comparisons.¹⁶ For example, nursing homes host relatively frail individuals with a life expectancy shorter than that of their age-equivalents who are able to live independently. Small area studies will therefore identify geographies with nursing homes as geographies with health concerns. Alternatively, this ecological bias will not affect interprovincial comparisons since all provinces provide nursing homes for their elder residents.

In this study, we try to present vital statistics in a format that is meaningful to policy makers by presenting small area maps of life expectancy. We pursue small area comparisons on a scale that is smaller than that currently available on a province-wide basis. We investigate the importance of socioeconomic factors and selective migration to nursing homes. In addition, we provide details of considerations and choices of data resources, functional geographic units, sources of error and statistical methods. These details are crucial to other small area applications that we have planned and will be helpful to researchers who want to pursue small area comparisons elsewhere in Canada.

Methods

Geographies

Nova Scotia comprises nine District Health Authorities, 18 counties, 52 census consolidated subdivisions, 110 census subdivisions, 1511 federal enumeration areas and 18,864 postal codes. Census subdivisions comprise cities, towns, villages, municipal districts and subdivisions of counties.¹⁷ A census consolidated subdivision is a grouping of census subdivisions where the smaller, more urban census subdivisions (towns, villages, etc.) are combined with the surrounding, larger, more rural census subdivision (municipal districts and subdivisions of counties).¹⁷ A census consolidated subdivision is a functional grouping: rural residents frequently have mailboxes in nearby communities with different postal codes, causing researchers to introduce misclassification when using geographic units smaller than census consolidated subdivisions.¹⁸ In addition, studies have reported that manual coding of place of residence inaccurately overcounts cities and undercounts their peripheral areas.¹⁹ For the above reasons we use “census consolidated subdivision” as our unit of comparison in non-urban geographies. The two urban areas, Metropolitan Halifax and Cape Breton Regional Municipality, are subdivided into geographies not exceeding 50,000 residents. Metropolitan Halifax is subdivided by grouping enumeration areas into 11 neighbourhoods, and the Cape Breton Regional Municipality is subdivided by grouping enumeration areas into four geographies largely divided by natural borders (lakes and rivers). This brings the number of customized geographies to a total of 64 with populations ranging from approximately 2,500 to 41,000 (see Appendix). The population of Nova Scotia is approximately 940,000.

Life expectancy and health deficiencies

We present estimates of life expectancy for each of the 64 geographies. Life expectancy is an easy-to-comprehend measure of local health status and thus preferable to health policy makers. We used Chiang's standard period life table methods to calculate life expectancy at birth and standard errors resulting from sampling error (SE_SE).²⁰ When calculating the combined life expectancy of females and males, we summed the radix and combined female and male deaths as a means to provide more stable estimates in geographies with small populations.²¹ All calculations are based upon 20 age categories (less than one year, one to four years, 17 consecutive five-year categories, and 90 years and older). Health deficiencies are defined as the difference between local life expectancy and that of the provincial average. This is broken down into cause-specific components, cardiovascular diseases (ICD 9: 390 to 459), cancer (ICD 9: 140 to 208), lung cancer (ICD 9: 162 and 163), colorectal cancer (ICD 9: 153 and 154), breast cancer (ICD 9: 174) and respiratory diseases (ICD 9: 460 to 519). These calculations are based upon cause-eliminated life table methods as described in detail elsewhere.^22,23

Population estimates

For the purpose of calculating life expectancy we need accurate population and mortality estimates for each of the 64 geographies. We considered three sources of population information: age and gender specific population counts for census years 1986, 1991, and 1996; Statistics Canada's online statistical database (CANSIM)²⁴ for 1986, 1991, and 1996; and population counts from the Nova Scotia Medical Services Insurance registration file at midyears that are available to us for 1996, 1997, 1998, and 1999. In 1996, for the province as a whole, the population estimates on the basis of the Medical Services Insurance registration file were 0.29% higher than on the basis of CANSIM and 2.73% higher than on the basis of the census. We adjusted the census population estimates in 1986 and 1991 for the underestimation observed in 1996. We estimated the population sizes for each age and gender subgroup in the years 1990 and 1992 to 1995 by applying cubic splines to the adjusted estimates for 1986 and 1991 and to the Medical Services Insurance counts for 1996 to 1999.²²

The postal code conversion file is a software instrument issued by Statistics Canada that allows researchers to allocate enumeration areas on the basis of postal codes.^18,25 In situations where a postal code constitutes more than one enumeration area, the postal code conversion file will select enumeration areas using a randomization procedure that considers the population size of the enumeration areas. As part of this study we investigated the variation in allocating postal coded places of residence to each of the 64 customized areas and the extent to which this affects the estimates of life expectancy. To do so, we repeated the allocation procedures 10 times and calculated standard errors of the 10 life expectancy estimates for each of the 64 geographies. Small standard errors are indicative of reproducible allocation procedures.

Mortality estimates

We obtained annual cause-specific mortality from Statistics Canada for the years 1995 to 1999, specified by age and gender. The available geographic information included postal code and geographic specifications by census subdivisions or county. Both postal code and geographic specifications introduce error in the allocation of mortality to geographic units. Postal codes require the use of the postal code conversion file (discussed above) and geographic specifications may cross boundaries or be missing and require random allocation. As part of the current study we assessed the extent of error introduced by allocating mortality when using postal codes, geographic specifications or the combination of both. We repeated the allocation procedures 10 times and subsequently calculated the standard error of the 10 repetitions. Small standard errors are indicative of reproducible allocation procedures.

Institutionalized population

Residents of nursing homes are relatively frail and have shorter life expectancies than their age-equivalents living independently. Migration of frail individuals from geographies without nursing homes to geographies with nursing homes will increase the life expectancy in the former and reduce the life expectancies in the latter. Small-area studies that aim to identify local health status and its determinants are hampered by this selective migration of frail individuals.¹⁶ Deaths in nursing homes are identifiable through their institutional postal codes. We retrieved the postal codes of the prior residential addressed for approximately 80% of these deaths. To illustrate the importance of the selective migration of frail individuals we compared life expectancy estimated using nursing home addresses with life expectancy estimated using previous residential addresses.

Determinants of Local Health Status

Area level measures of socioeconomic status, average household income and unemployment rate are investigated with respect to their association with local life expectancy. This information is taken from the 1996 Canada Census.

Statistical Methods

We estimated life expectancy and health deficiencies over the calendar period of 1995 to 1999 for each of the 64 geographies. These estimates fluctuate more than expected on the basis of sampling error (overdispersion) as a result of the sparseness of (cause specific) mortality and varying population sizes of the 64 geographies.^8,13 If overdispersion is ignored it creates the impression of spurious geographic variation and consequent instable estimates of associations with covariates such as areal socioeconomic characteristics.⁸ Bayesian hierarchical or multilevel models have been suggested as appropriate methods to analyze such small area data.⁸ We considered a multilevel model whereby information of the 64 geographies and their direct neighboring geographies are pooled (level 1) resulting in robust estimates of the geography-specific life expectancy (level 2). This model also allows us to incorporate the various sources of standard error described above. We will refer to the empirical Bayes estimates generated by this model as “spatially smoothed estimates” in the remainder of this manuscript.

Using multilevel approaches, we further considered geographies (level 1) within regions (level 2) to generate empirical Bayes (spatially smoothed) estimates of regional variation. The four regions considered include: non-metropolitan mainland, metropolitan Halifax, non-metropolitan Cape Breton Island, and Cape Breton Regional Municipality (see appendix). With this multilevel model we also analyzed the association of socioeconomic characteristics, income and unemployment rate to life expectancy. The analyses were conducted by using HLM5 and S-PLUS 2000.

Results

Health deficiencies, defined as the differences between local life expectancy and the provincial average, are visualized in Figure 1 through colour patterns with red, indicative of reduced life expectancy and blue, indicative of prolonged life expectancy. Various geographies have substantially reduced, dark red, or substantially prolonged life expectancy, dark blue (Figure 1 top panel). Life expectancies are also listed in the appendix along with the population size, numbers of deaths and various sources of standard error. Sampling error (SE_SE) is determined by the number of residents and deaths in each age and gender subgroup. Geographies with large populations have generally small SE_SE (appendix). Standard error resulting from the allocation procedures in estimating the population sizes, SE_POP, is negligible relative to SE_SE. The appendix lists three estimates of standard error corresponding with three means of allocating mortality: SE_M1, if only geographic specifications are considered, is substantial relative to SE_SE in urban areas. SE_M2 considers only postal code information and is substantial particularly in rural areas. SE_M3 uses both geographic and postal code information and is generally less than SE_M1 and SE_M2, but for a few geographies still substantial relative to SE_SE.

Both SE_SE and SE_M3 were considered in the multilevel model generating the spatially smoothed estimates that are listed in the appendix and depicted in the bottom panel of Figure 1. These smoothed estimates allow better judgment of the geographic distribution: Cape Breton Island has reduced life expectancy and within the Halifax Regional Municipality considerable variation persists. The differences between the crude and spatially smoothed estimates are frequently large in geographies with small populations (Figure 2 – top panel). The differences of estimates with and without adjustment for selective migration of nursing home residents is of a smaller magnitude and not as clearly related to the population size of the geography (Figure 2 – bottom panel).

Health deficiencies resulting from cardiovascular and cancer mortality are displayed in Figure 3 and reveal distinct geographic patterns.

The health concerns in Cape Breton County have previously been addressed in ecological studies.^{16,22,23,27,28} Suggested underlying factors include life style choices such as smoking and obesity, participation in screening programs, environmental conditions and socioeconomic factors.^{22,23,29–31} This study provides additional understanding by demonstrating that regional differences are only in part explained by income or unemployment. Health differences elsewhere in the province have not been evaluated previously. This study provides a province-wide evaluation and reveals that Cape Breton County is not the only area with major health concerns. The “North End” neighbourhood in Halifax exhibits health concerns of similar magnitude in contrast to the wealth and health of other neighbourhoods.

Table 1 presents the univariate association of income and unemployment with life expectancy. An increment in income of $10,000 is associated with an increase of 0.956 years in life expectancy, whereas an increment in unemployment of 10% is associated with a reduction of 0.862 years in life expectancy. Multilevel regression (spatially smoothed) estimates an increment in income of $10,000 to be associated with an increase of 0.617 years in life expectancy and the association with unemployment not to be statistically significant. Table 2 presents the extent of regional variation: Life expectancy in the Cape Breton Regional Municipality is estimated to be 1.46 less than in the non-metropolitan mainland, and 1.57 (in Table 2: 1.46+0.11) less than in the Halifax Regional Municipality. These differences are statistically significant and only partly explained by differences in income or unemployment (Table 2). Table 3 shows the regional variation in life expectancy broken down in disease specific components. Due to cardiovascular mortality, life expectancy of female residents of Cape Breton Regional Municipality is 0.36 years less than that in the non-metropolitan mainland; for male residents it is 0.74 years. The estimates are higher for cardiovascular disease than for cancer.

Discussion

This study demonstrates the geographic variation of health within Nova Scotia. Geographies with major health concerns include Cape Breton Regional Municipality and the disadvantaged “North End” neighbourhood within metropolitan Halifax. Although socioeconomic factors are important determinants of health, they are only partially responsible for the observed variation in health. This study also demonstrates the feasibility of small area comparisons of health and the importance of appropriate allocation procedures, statistical methods and selective migration.

We presented small area maps visualizing geographic patterns in health and identifying geographies and regions of concern. Life expectancies in some geographies are reduced by more than 1.5 years relative to the provincial average which, in turn, is approximately one year less than the Canadian average.²⁶ These differences are substantial: health in these geographies lag 10 to 15 years behind that of Canadians, if one considers the national increase in life expectancy of approximately two years per decade.²²

The aim of this study was to pursue small area comparisons at a scale that is smaller than what is currently available at a provincial level. We demonstrated the feasibility of mapping postal code data at the level of census-consolidated subdivisions in rural areas and neighbourhoods in urban areas. The various challenges we experienced in mapping at this small area level, including allocation procedures, statistical methods and selective migration, are discussed below:

Allocation procedures. The postal code conversion file has become a crucial instrument for the allocation of postal code information to geographic locations and has enabled the conduct of various research.^3,18,32 Its appropriateness in small area studies has been evaluated here. For the allocation of large counts, such as for whole residential populations, the reproducibility of the postal code conversion file appeared excellent and was reflected in very small standard errors. However, if counts, such as mortality counts, are sparse, the reproducibility decreases. Investigators using the postal code conversion file should be aware of this additional source of error. In this study we demonstrated that this error is particularly present in rural areas, and decreases when considering additional geographic specifications. To reduce this error, investigators may alternatively want to repeat the allocations with the postal code conversion file and consider the average value of the repeated allocations.

Statistical methods. We have presented maps of both crude and spatially smoothed estimates of life expectancy. Interpretation of these maps warrants caution: the crude estimates exhibit overdispersion and therefore create the impression of spurious geographic variation. The spatially smoothed estimates are a solution to overdispersion and are thus most appropriate for the judgment of geographic variation. All smoothing procedures, including those used in this study, are to some extent arbitrary.⁸ We chose to consider the life expectancy of all neighbouring geographies in the smoothing procedures, although others investigators may have chosen differently.^13,14 These choices affect the magnitude of the smoothed estimates of life expectancy and should be considered when judging geographic variation. The importance of the choice of statistical methods was also reflected in the analyses of socioeconomic factors, where crude and spatially smoothed estimates differed considerably.

Selective migration of healthy or frail subgroups can affect local estimates of life expectancy and thus ecological comparisons, particularly small area comparisons.¹⁶ In this study we demonstrated that in five of the 64 geographies (7.8%) selective migration to nursing homes altered local estimates of life expectancy by more than one year. Clearly, and in addition to established and causal risk factors, the presence of nursing homes should also be considered as a factor affecting estimates of life expectancy of small areas. Since the analysis of selective migration was based on the incomplete prior residential addresses of nursing home residents, the actual effect of selective migration is likely larger.

We have aimed to present health data in an easily comprehended manner and have chosen the format of provincial maps of life expectancy. They revealed various matters that are important to policy makers, such as the revelation that the Cape Breton Regional Municipality is not the only area with major health concerns. In addition, they revealed distinct geographic patterns in the underlying causes of death. In this respect, cardiovascular disease is demonstrated to be the single most important cause of death responsible for the health deficiencies in the Cape Breton Regional Municipality, whereas past discussions have primarily focused on the high cancer rates in this region.^22,27,28 In addition to the well established relationship between wealth and health, policy makers are now informed that neither income nor unemployment explain the provincial health disparities and the health concerns of Cape Breton Regional Municipality. These are examples of how small area comparisons may contribute to decision processes of policy makers. More applications are to be expected from future small area comparisons of morbidity, health care utilization and other descriptive measures and determinants of health.

Acknowledgments

Support for this study is provided through funding by the Canada Foundation for Innovation, the Dalhousie Medical Research Foundation, the Nova Scotia Health Research Foundation and a Canadian Institutes of Health Research Career Award to Dr. Veugelers.

The authors thank David Elliott, Angela Fitzgerald, Michael Pennock, Chris Skedgel, Mark Smith and Alexandra Yip for their helpful assistance.

References

1.    Rose G. The Strategy of Preventive Medicine. Oxford University Press. Oxford, 1992.

2.    Ross NA, Wolfson MC, Dunn JR, Berthelot J-M, Kaplan GA, Lynch JW. Relation between income inequality and mortality in Canada and in the United States: cross sectional assessment using census data and vital statistics. BMJ 2000; 320: 898–902.

3.    Veugelers PJ, Yip AM, Kephart G. Proximal and Contextual Socioeconomic Determinants of Mortality: Multilevel Approaches in a Setting with Universal Health Care Coverage. Am J Epidemiol 2001; 154:725–732.

4.    Province of Nova Scotia. The Course Ahead: for the fiscal year 2000-01. Government of Nova Scotia, 2000.

5.    Susser M. Does Risk Factor Epidemiology Put Epidemiology at Risk? Peering into the Future. J Epidemiol Community Health 1998; 52:608–611.

6.    McMichael AJ. Prisoners of the Proximate: Loosening the Constraints on Epidemiology in an Age of Change. Am J Epidemiol 1999;149:887–897.

7.    Olsen S, Martuzzi M, Elliott P. Cluster Analysis and Disease Mapping – Why, When and How? A Step by Step Guide. BMJ 1996; 313:863–866.

8.    Lawson A, Biggeri A, Böhning D, Lesaffre E, Viel J-F, Bertollini R. Disease Mapping and Risk Assessment for Public Health. John Wiley & Sons. Toronto 1999.

9.    Bertollini R, Martuzzi M. Disease Mapping and Public Health Decision-Making: Report of a WHO Meeting. Am J Public Health 1999;89:780.

10.    Elliott P, Cuzick J, English D, Stern R. Geographical and Environmental Epidemiology. Methods for Small-Area Studies. Oxford University Press. New York, 1997.

11.    Manuel DG, Goel V, Williams JI, Corey P. Health-adjusted Life Expectancy at the Local Level in Ontario. Chronic Dis Can 2000; 21:73–80.

12.    Wilkins R. Health Expectancy by Local Area in Montreal: a Summary of Findings. Can J Public Health 1986;77:216–222.

13.    Clayton D, Kaldor J. Empirical Bayes Estimates of Age-standardized Relative Risks for the Use in Disease Mapping. Biometrics 1987;43:671–681.

14.    Martuzzi M, Elliott P. Empirical Bayes Estimation of Small Area Prevalence of Non-Rare Conditions. Statist Med 1996:15;1867–1873.

15.    Wakefield J, Elliott P. Issues in the Statistical Analysis of Small Area Health Data. Statist Med 1999;18:2377–2399.

16.    Veugelers PJ, Guernsey JR. Sensitivity analysis of selective migration in ecologic comparisons of health. Epidemiology 1999;10: 784–785.

17.    Statistics Canada. 1996 Census Dictionary: Cat. No. 92-351-UIE

18.    Wilkins R. Automated Geographic Coding Based on the Statistics Canada Postal Code Conversion Files. August 2001. Statistics Canada Cat. No. 82F0086-XDB.

19.    Manuel DG, Goel V, Williams JI. The Derivation of Life Tables for Local Areas. Chronic Dis Can 1998; 19:52–56.

20.    Chiang CL. The Life Table and Its Construction. Introduction to Stochastic Process in Biostatistics. New York: John Wiley & Sons, Inc. 1968.

21.    Manton KG, Stallard E. Chronic Disease Modeling: Measurement and Evaluation of the Risks of Chronic Disease Processes. New York: Oxford University Press. 1988.

22.    Veugelers PJ, Guernsey JR. Health deficiencies in Cape Breton County, Nova Scotia, Canada, 1950–1995. Epidemiology 1999; 10:495–499.

23.    Veugelers PJ, Kim AL, Guernsey JR. Inequalities in Health. Analytic Approaches based on Life Expectancy and suitable for Small Area Comparisons. J Epidemiol Community Health 2000;54:375–380.

24.    Statistics Canada website: http://www. statistics.ca

25.    Wilkins R. Geocodes/PCCF Version 2 User's Guide. Automated Geographic Coding Based on the Statistics Canada Postal Code Conversion File. Health Statistics Division, Statistics Canada, Ottawa, July 1997.

26.    Millar WJ. Life expectancy of Canadians. Health Rep 1995;7(3): 23–26.

27.    Mao Y, Morrison H, Semenciw R. Mortality in Cape Breton, Nova Scotia, 1971–1983. Chronic Disease in Canada Special Report No. 11. December, 1985, Health and Welfare, Canada.

28.    Guernsey JR, Dewar R, Weerasinghe S, Kirkland SA, Veugelers PJ. Incidence of Cancer in Sydney and Cape Breton County, Nova Scotia 1979–1997. Can J Public Health 2000, 91:285–292.

29.    Fitzgerald AL, Veugelers PJ, MacLean DR. Dietary reference intakes: a comparison with dietary intake in Nova Scotia. Can J Dietetic Practice and Research. November 2001, accepted for publication.

30.    Nova Scotia Department of Health. Smoking Ban in Public Places. Public Opinion Survey. Nova Scotia Department of Health, 1995.

31.    Public Affairs Department Cape Breton District Health Authority. Our Health. Cape Breton District Health Authority, 2001.

32.    Demissie K, Hanley JA, Menzies D, Joseph L, Ernst P. Agreement in Measuring Socio-economic Status: Area-based versus Individual Measures. Chronic Dis Can 2000; 21:1–7.

Author References

Paul J Veugelers and Shane Hornibrook, Department of Community Health and Epidemiology, Faculty of Medicine, Dalhousie Univeristy

Correspondence: Paul J Veugelers, Department of Community Health and Epidemiology, Dalhousie University, 5849 University Avenue, Halifax, Nova Scotia, Canada B3H 4H7; Fax: (902) 494-1597; E-mail: paul.veugelers@dal.ca

Appendix

[Previous] [Table of Contents] [Next]