International Journal of Health

Background: The aims of this study were to determine whether observed geographic variations in breast cancer incidence are random or statistically significant, whether statistically significant excesses are temporary or time-persistent, and whether they can be explained by covariates such as socioeconomic status (SES) or urban/rural status?


Background
The majority of work in Australia describing the association between the health status, use of health services and socioeconomic status of the population uses an areabased measure of socioeconomic status. It is necessary to use a proxy measure (the socioeconomic status of the population in the area) because there is no direct measure in the major administrative health record collections of the socioeconomic status of the individual about whom the event is recorded.
However, the application of an area-based measure requires a number of assumptions, including that people who move do so between, or within, geographic areas of similar socioeconomic status; and that the (often large) areas used in these analyses provide a reliable indication of the socioeconomic status and health service utilisation of the individuals in the area about whom the event is recorded. Area level socioeconomic status can also be considered as an independent predictor. For example, an individual with low socioeconomic status in an area of higher socioeconomic status is more likely to have better health outcomes than their counterpart in an area of lower socioeconomic status [1,2]. This aspect is not addressed in this paper.
In relation to this latter point, Hyndman et al [3] found that "Misclassification of individuals to SES groups based on the basis of postcode caused an underestimation of the true relationship between SES and health-related measures. A reduction of this misclassification by using smaller spatial areas, such as CD or census enumeration districts, will provide improved validity in estimating the true relationship." A reduction in strength of correlation with increasing size of area is consistent with the results of this paper. In a study of hospitalisations in Michigan, USA, Hofer et al [4] found that 'the impact of socioeconomic characteristics on hospitalization rates is consistent when measured by individual or community-level measures'. This is an encouraging finding for those limited to using area-based data.
Another limitation of the majority of Australian healthrelated datasets is that they record events (eg., hospital inpatient separations, services by general medical practitioners), rather than individuals.
The analysis in this report uses the Western Australian Data Linkage System to explore the reliability of area data as a proxy for socioeconomic disadvantage when analysed for the relatively large geographic units often used in health-related research: it also addresses the limitations of using data about events, rather than individuals. It does this by examining the extent to which hospital inpatient separation rates vary, both overall and by socioeconomic status of area of residence, when calculated at various levels of aggregation, from Census Collection District (CD)the smallest area level for which a measure of socioeconomic status is available -to the larger units of postcode and Statistical Local Area (SLA). Methods applied include the calculation of correlation coefficients and examination of hospital separation rates by quintile of socioeconomic disadvantage of area, separately for events and individuals.
The report also examines the reliability of the socioeconomic status of the address over time, by examining the extent of change in socioeconomic status of area of residence for individuals with repeat hospital episodes over a five year period.
The analysis shows that aggregating data to larger area reduces the gap between the index scores for the most disadvantaged and least disadvantaged areas, with the greatest impact on the scores for the most disadvantaged areas. This results in an understatement of the extent of disadvantage in the most disadvantaged areas, as well as an understatement in inequality between the most well off and the poorest areas.

Individuals
Over the five years from 1994 to 1998, a total of 358 948 residents of Perth were admitted to a hospital in Western Australia on one or more occasion, an average of 71 750 individuals admitted per annum. Just over half (53.6%) the individuals admitted were females; 46.4% were males.
The rate of individuals admitted was 16.4% higher for females (247.6 separations per 1000 population) than for males, (212.7 separations per 1000 population) ( Table  1). As can be seen in Figure 1, the rates of males and females admitted vary notably by age. For females, the highest rate is in the 30 to 34 year age group (with a further three of the five highest female rates between ages 20 and 39 years), with the second highest rate in the 80 years and over age group. The highest male rate, in the 80 years and over age group, is substantially above the next highest male rates, in the 50 to 69 year age groups.
A total of 358 768 Perth residents had one admission to a Western Australian hospital over the five years from 1994 to 1998, with a further 298 805 people admitted on two or more occasions ( Table 2). The number of people with two or more admissions in any period is higher in the earlier years, as the more time that passes the greater the opportunity for a second admission. That is, those with a first admission in 1994 have had more time to record a second admission than have those with a first admission in 1995: thus the greater number with two or more admissions in 1994.
Just over half (54.6%) those admitted to hospital had one admission over this period, and more than one third (36.0%) had between two and four admissions, together comprising 90.6% of those admitted (Table 3).
Females accounted for just over half (53.6%) of those admitted once, compared with 59.7% of those admitted more than once. For males, the proportions were 46.4% and 40.3%, respectively.

Separations
There were 1 665 308 separations of Perth residents from Western Australian hospitals, an average of 2.53 separations per person admitted over the five years from 1994 to   Perth residents admitted to hospital, by age and sex at first admission, 1994-98 Figure 1 Perth residents admitted to hospital, by age and sex at first admission, 1994-98. Perth population is at 30 June 1996. Per cent shown is of males and females separately, not for persons.
1998. Over half (55.1%) of the separations were of females and 44.9% were of males. Figure 2 shows the profiles of males and females, by age, for both individuals admitted (as in Figure 1) and separations. For males, the proportion of individuals admitted is highest at ages 20 to 49 years, dropping away at younger and older ages, with the latter exhibiting a particularly marked drop. Total separations for males are generally highest at older ages (the highest at ages 70 to 74 years), reflecting the higher number of separations per person. The notable exception is the high proportion of separations in the 0 to 4 year age group. The profile of the proportion of females admitted is similar to that for males, although it is somewhat distended at ages 20 to 39 years. The proportion of separations of females at ages 25 to 54 years closely follows that for females (individuals) admitted.
The main differences in the profiles of male and female separations are evident at the youngest ages (higher proportions of males), from ages 20 to 44 years (higher proportions of females) and from 50 to 79 years (higher proportions of males). The ages at which the highest rates of admissions of individuals and of multiple admissions (the gap between the separations and admitted profiles) occur are clearly visible in the chart.
Unlike the rates for individuals admitted (Table 1, above), the highest rates for separations of both males and females occur in the oldest age groups ( Table 4). The five highest rates for both males and females are in the age groups 60 to 64 years and over, with male rates higher (and often substantially so) than female rates. Also of note is the high rate of separations for females at ages 30 to 34 years (1,672.2 admissions per 1000 population): this is the sixth highest rate for females, and is more than twice the rate for males at the same age (729.5 separations per 1000 population).

Effect of aggregation of areas on disadvantage scores
As noted, the majority of the analysis by socioeconomic status undertaken in the health sector in Australia is area based, and uses the postcode or SLA as the unit of analysis. This raises the question of the extent to which area based analyses at the postcode or SLA level provide a reliable indication of the socioeconomic status and health service utilisation of the individuals admitted. This report Perth residents admitted to hospital and total separations, by age and sex, 1994-98 Figure 2 Perth residents admitted to hospital and total separations, by age and sex, 1994-98. Perth population is at 30 June 1996. Per cent shown is of males and females separately, not for persons. explores the reliability of postcode or SLA level data by examining the extent to which rates of individuals admitted and separations vary when calculated at various levels of aggregation (CD, postcode and SLA). Ideally, the comparison would be between the socioeconomic status of individuals and of areas; however, the smallest area level for which a measure of socioeconomic status is available is the CD.
Variation in the minimum and maximum Index of Relative Socio-economic Disadvantage (IRSD) scores when calculated at the CD, postcode and SLA level is striking and clearly shows the value of the smaller unit in area based analyses ( Table 5). The range at the CD level is from a minimum index score of 532 to a maximum index score of 1221, a differential of 2.3 times. When individuals and separations are analysed by postcode, the range in the IRSD scores is narrower, from 863 to 1168 (a differential of 1.4). At the SLA level it is slightly lower again (a differential of 1.3). The effect of aggregation to the larger areas is most noticeable in the minimum IRSD score, increasing the minimum score by 70.5% from the CD level to the SLA level. At the other end of the scale, the maximum score varies little, dropping by 4.0%. That is, the greatest loss in specificity in the IRSD score is in the most disadvantaged areas.
Thus, the use of larger area aggregates reduces the gap between the index scores for the most disadvantaged and least disadvantaged areas (thus lessening the extent of inequality between these areas), with the greatest impact on the scores for the most disadvantaged areas (thus understating the extent of inequality in these areas). Notably, the difference between the maximum and minimum scores, and the absolute level of the scores, is much less marked between the postcode and SLA.
There was a strong association between the IRSD scores for CDs and those for postcode of usual address at the first admission (a Spearman correlation coefficient of 0.74). The correlations were between CDs grouped to quintiles and postcodes grouped to quintiles, ranked by the IRSD, and not between individual CDs and postcodes. A weaker association was found between the quintiles for CDs and those for SLAs (0.64 for people with one separation and 0.63 for people with more than one separation) ( Table 6). There was little difference in correlation coefficients for those who had moved address. Similar Spearman correlation coefficients were calculated for raw IRSD scores.

Effect of aggregation of areas on separation rates
Data at the CD level for the five years from 1994 to 1998 show a variation in rates of individuals admitted from 51 442 admissions per 100 000 population in the most advantaged areas to 58 130 admissions per 100 000 population in the most disadvantaged areas (Table 7). This is a differential of 13%. The differential in separation rates is substantially higher, at 47%, reflecting multiple admissions.
When data are aggregated to postcode area or SLA, the differentials in separation rates between Quintile 5 and  Quintile 1 areas are smaller (differentials of 1.23 and 1.20, respectively) than at the CD level (a differential of 1.47) ( Table 7). In the case of postcodes, this is largely because of the lower separation rate in Quintile 5 areas (likely to be a result of the process of aggregating CDs), whereas for SLAs it is a combination of a lower separation rate in Quintile 5 areas and a higher rate in Quintile 1 areas (likely to be a result of the aggregation process, exacerbated by the variable size of SLAs -see section titled 'Methods, Area' under 'Methods.' The differential in rates of individuals admitted is the same for data at the SLA and CD level, but higher for postcode areas. These results again reflect the difficulty inherent in producing groups of approximately equal populations. While just over half (54.6%) those admitted to hospital had one separation over this period, the proportion varied from 56.3% in Quintile 1 to 51.9% in Quintile 5 (Table  8). This is as expected, with people from the most disadvantaged areas representing a smaller proportion of The extent of any inequality is shown by the rate ratio, which expresses the ratio of the rate in Quintile 5 to the rate in Quintile 1; rate ratios indicating differing significantly from 1.0 are shown with * p < 0.05; ** p < 0.01; *** p < 0.001.   Table 9).
The average number of admissions per person for people admitted to hospital on more than one occasion over the five years to 1998 was 4.4; this varied from 4.2 separations per person admitted in the least disadvantaged areas to 4.7 in the most disadvantaged areas.

Reliability over time of address as a proxy for socioeconomic status
Studies using the address of usual residence as a proxy for socioeconomic status require two important assumptions. They are that: • people who move do so within, or between, areas of similar socioeconomic status; and that • the areas used in an area based analysis (which can vary in size and are quite often large) provide a reliable indication as to the socioeconomic status and use of health services of the individuals in the area.
Data from the 1996 Census show that 53.5% of Perth's population at the 1996 Census reported that they had a different address to that at the previous Census, five years earlier [5]. Data were not available to compare the IRSD of the first and last SLA of address of the Perth population who moved. However, almost one quarter (24.0%) of Perth residents who moved between the 1991 and 1996 Censuses moved to an address within the same SLA. That is, some 59.3% of the population were in the same SLA after five years (either moved within the SLA, or did not move). This is an encouraging statistic for area based analyses.
Similarly, almost four out of five people admitted to hospital more than once in a five year period had not moved (out of the CD of their address at the first separation) by the time of their second separation. For example, of the 298 809 people admitted to a Perth hospital more than once over the five year period 1994 to 1998, over three quarters (78.6%, 64 075 people) had the same address at the time of the second separation. People were recorded as having 'moved' if the CD of their address changed between the first and last separation over the period from 1994 to 1998. Movement to a different address within a CD was not included.
The following table illustrates, for people with multiple admissions, the extent of movement by socioeconomic status. For this part of the analysis, the CD of first and last separation have been allocated to quintiles of socioeconomic disadvantage of area, to provide a comparison of the extent of movement between different levels of socioeconomic status. The construction of the quintiles is described in the section titled 'Methods, Measurement of socioeconomic status' under 'Methods.' Table 10 shows, for people who moved to an address in another CD, that: • people from the most well off areas are less likely to have moved to areas of greatly different socioeconomic status The extent of any inequality is shown by the rate ratio, which expresses the ratio of the rate in Quintile 5 to the rate in Quintile 1; rate ratios differing significantly from 1.0 are shown with * p < 0.05; ** p < 0.01; *** p < 0.001.
(ie, changed quintiles) than are those from the most disadvantaged areas -40.2% of people in the most advantaged areas (Quintile 1) remained there, despite moving from the CD of their first separation. The proportion in the most disadvantaged (Quintile 5) areas was a lower 30.5%; • while there is movement right across the socioeconomic profile, most movement is between adjacent quintiles. For example, of the 18 875 people who lived in the most disadvantaged areas at their first separation (and moved before a subsequent admission), 71.2% had moved to a CD in the same or next ranked quintile (Quintiles 5 or 4), with just 4.6% moving to the most advantaged areas. Similarly, of the 9 537 people in the most well off areas at their first separation, 63.0% had moved to a CD in the same or next ranked quintile (Quintiles 1 or 2), with a similarly low proportion (4.7%) moving to the most disadvantaged areas; • the most substantial movement between quintiles was of people moving from an address rated as Quintile 5 to one rated as Quintile 4 (40.7%); this was marginally higher than the proportions moving within Quintiles 4 or 1 (40.3% and 40.2%, respectively).
There is a strong association between the quintile of socioeconomic disadvantage of area at the first and the last discharge when analysed by CD (a correlation coefficient of 0.88) or SLA (a correlation coefficient of 0.89) of usual address (Table 11). This supports the earlier finding that people admitted to hospital who had moved between episodes, moved to or within areas of similar socioeconomic status. The weaker correlations between CD and SLA (see table) highlight the loss in specificity of the index score when aggregated to the (larger) SLA level.

Conclusions
The analysis shows that, for Perth residents admitted to hospital, the use of larger area aggregates reduces the gap between the index scores for the most disadvantaged and least disadvantaged areas, thus understating the extent of inequality between these areas. The greatest impact of aggregation of areas is on the scores for the most disadvantaged areas. This results in an understatement of the extent of disadvantage in the most disadvantaged areas, as well as an understatement in the extent of inequality between the most well off and the poorest areas.
Further, the analysis shows that a more people from the most disadvantaged areas are admitted to hospital than  from the most well-off areas (13% more), and that these people have more separations overall (47% more), as a result of larger numbers of multiple admissions.
As regards the extent of movement, four out of five people admitted to hospital more than once in a five year period had not moved (out of the CD of their address at the first separation) by the time of their second separation. In addition: • people from the most well off areas are less likely to have moved to areas of greatly different socioeconomic status than are those from the most disadvantaged areas; • while there is movement right across the socioeconomic profile, most movement out of a quintile is to areas in adjacent quintiles; and • the most substantial movement between quintiles was of people moving from an address rated as Quintile 5 to an address rated as Quintile 4, although this was only marginally higher than the proportions moving within Quintiles 4 or 1.
In summary, postcode level and SLA level data provide a reliable indication of socioeconomic disadvantage of area, when compared with CD-level data. That is, the association between rates of total separations and individuals admitted and socioeconomic disadvantage of area evident at the smallest area level (CD) is also evident in the higher level area aggregates of postcode and SLA.
It is reasonable to assume that similar relationships exist in other Australian cities, as well as in other health-related activity (eg. visits to general medical practitioners).
Given the widespread use in Australia of area based analyses at the postcode and SLA level, and the limitations of CDs an area level for the analysis of most health datasets, it is important to know that such analyses provide a reliable indication of the direction and underlying strength of the influence of socioeconomic factors in hospital admissions rates.
This is not to imply that the postcode or SLA are the ideal areal unit for analysis, nor that data for Collection District would be. The ideal population size for area-level analysis is likely to vary dependent on the number of cases in the dataset under analysis. For datasets with a large number of cases per capita (eg. services by general medical practitioners) the number will be smaller than those with a small number of cases per capita (eg. deaths), even with aggregation of data over a number of years. May SLAs have much larger populations than are necessary to produce reliable results; and the populations of most CDs are too small (see Table 12). HealthWIZ [6], the National Social Health Database, comprising among the most widely available small area datasets in Australia, seeks to provide health service use and health status data for areas with populations of approximately 10 000. This is a useful benchmark.
It is also clear that data as to socioeconomic position at the smallest area level possible or, more importantly, of individuals, would also be of value. Were data to be available across Australia on a similar basis to that from the Western Australian Data Linkage System, it would be possible to undertake research of greater policy-relevance than is currently possible with the existing separationsbased national database. Such moves are under consideration in several Australian States.
Further, linking data (eg, using probabilistic linkage) for individuals in the Western Australian Data Linkage System to the Australian Bureau of Statistics Population Census has the potential to add considerable value to such analyses. For example, it would be possible to examine an individual's characteristics of education, occupation, labour force status, housing tenure etc., and to more directly examine the relationships between the number of individuals admitted and total separations and these important socioeconomic variables. Linkage to death registration data would also be valuable in understanding more about outcomes related to socioeconomic status. This latter example is a possibility under recently announced plans for the ABS to test the linking of 2006 Census of Housing and Population data to other datasets, such as deaths registrations, held under their Act. This is similar to the approach elsewhere, including New Zealand [7]. It is to be hoped that such arrangements can be put in place in Australia in the near future.

Terminology
The report addresses differences in the number of individuals admitted and the number of separations they incurred. These are described as 'individuals', or individu- als admitted' and separations (the total number of separations, where an individual may have had one or more episodes of hospitalisation over the period of the analysis). 'Separation' is the term describing a completed hospital episode: it is defined in the section titled 'Glossary, Separation' under 'Glossary.'

Data sources
Details of all separations to public and private hospitals in Western Australia for the five years from 1994 to 1998 were extracted from the Western Australian Hospital Morbidity Database (HMDS). Any separation records thought to belong to the same person had previously been linked together within the Data Linkage System, permitting analyses to be performed for both separations and individual persons. The population used in calculating rates is the 1996 Census population.
The analysis has been limited to separations of residents of Perth, but includes separations occurring at any public acute or private hospitals in Western Australia. The HMDS includes address details for each separation from a hospital in Western Australia since 1993. These addresses have been linked to a Western Australian street address database to assign northing and easting points (geo-codes). These points were then assigned to the appropriate 1991 or 1996 CD using the ABS CData96 mapping tool. The postcode and SLA of the address were then determined by allocation of CDs to postcode or SLA. The boundaries for CDs and SLAs are consistent. However, boundaries for CDs and postcodes are not, so CDs were allocated to postcodes on a 'best fit' basis (see Glossary).

Area
Consequently, comparisons can be made between results for CDs and postcode areas, CDs and SLAs and postcode areas and SLAs. This is particularly important, as much of the area analysis undertaken in the health sector in Australia uses the postcode or the SLA, as a majority of data are only available at these area levels, and it is widely accepted that the larger the area, the less homogenous the population is likely to be.
There were 2 297 CDs in Perth at the 1996 Census, with 105 postcodes and 37 SLAs. The average population size at each of these area levels is shown in Table 12; these data emphasise the variation in size of the areas at each area level. Each area level (CD, postcode or SLA) was allocated to one of five groups (quintiles). For example, for SLAs, Quintile 1 comprises the SLAs with the highest IRSD scores (most advantaged areas), and Quintile 5 comprises the SLAs with the lowest IRSD score (most disadvantaged areas): each quintile comprises approximately 20% of the Perth population. This process does not provide an exact allocation of population, so the resultant populations are only 'approximately' equal, and the larger the areal units being allocated, the less likely they are to be equal. As shown in Table 13, when areas were ranked by their IRSD score at the CD level and then grouped to produce quintiles, the resultant populations were relatively close to the ideal population in each quintile of 245 607 (one fifth of 1 288 036). The quintiles based on postcode areas had rather 'lumpier' populations (greater variation around the one fifth figure of 254 859 per quintile -and a higher total of 1 274 297, due to boundary differences between CDs and postcodes. The quintiles based on SLAs were the most variable. For example, the SLA of Wanneroo -South West (with a population of 103 176) had a score marginally below the cut-off score between Quintile 1 and Quintile 2. However, the inclusion of Wanneroo -South West in Quintile 2 resulted in popula-

Analysis
Three (different) IRSD scores were added to each hospital separation record, based on the CD, postcode or SLA that had been previously assigned to the address on that record. It should be noted that these IRSD scores were actually the average score for the particular CD, postcode or SLA as calculated from 1996 Census data. Quintile ranks for each aggregation level were also applied using population weighting as described above.
For analyses involving multiple admissions, the IRSD value used was that for the first separation in the five-year period. These 'first' separations were isolated using the internal links between separation records for the same person and the separation date. Of course, many of these 'first' separations could have been preceded by separations occurring before 1994.
Rates are crude rates, per 100 000 population. Ideally the data would have been standardised (by the indirect method). However, access to the source data were limited and to requested tables, and standardisation was not an option.
As the data were from a complete enumeration (all admissions to hospital), confidence intervals were only calculated for measures of difference (in this case, rate ratios).
The Spearman Rank Correlation has been used in the analysis to indicate the degree of correlation between pairs of variables.

CD
The Collection District (CD) is the smallest area level in the Australian Bureau of Statistics' statistical geography and is primarily an area used in the five yearly population census.

Index of Relative Socio-Economic Disadvantage
The Index of Relative Socio-Economic Disadvantage (IRSD) is one of five Socio-Economic Indexes for Areas produced by the Australian Bureau of Statistics at recent population censuses. Produced using Principal Components Analysis, it summarises information available from variables related to education, occupation, income, family structure, race (the proportion of Indigenous people), ethnicity (poor proficiency in use of the English language) and housing. The variables are expressed as percentages of the relevant population. The IRSD is available at the Census Collection District level and was then be calculated for postcodes and SLAs by weighting the CD level scores by their population. The IRSD is calculated to show the relativity of areas to the Australian average for the particular set of variables which comprise it; this average score is set at 1000. Scores below 1000 indicate areas with relative disadvantaged populations under this measure, and scores above 1000 indicate areas with relatively advantaged populations. The IRSD scores at the Census Collection District (CD) level have been grouped to postal area, an area developed by ABS for the presentation of population counts and other Census data from the fiveyearly population censuses to approximate postcode areas, as the ABS does not collect the postcode at the Census.

Separation
The term describing a completed hospital episode is a 'separation'. At the time of admission to hospital, the age, sex, address of usual residence and other personal details of the patient are recorded. At the end of the episode, at the time of separation from hospital, details of the episode itself are recorded, including the date, time and method of separation (discharge, death or transfer of a patient to another care setting eg. hospital, nursing home). Consequently, hospital inpatient data collections are based on separations.

Postal area
The postal area is an area developed by ABS for the presentation of population counts and other Census data from the five-yearly population censuses. It approximates postcode areas, as the ABS does not collect the Australia Post postcode at the Census. Postal areas comprise Census Collection Districts (CDs) grouped to approximate postcode areas. Where a CD does not fit entirely within a postcode area, it is allocated to the postcode area into which the population largely falls. Where a CD covers more than one postcode area, the total CD population is allocated to one postcode.
The IRSD scores at the Census Collection District (CD) level have been allocated to postal areas as described in the section titled 'Methods, Index of Relative Socio-Economic Disadvantage' under 'Methods.'. Similarly, the postal area of each separation was approximated from the CD of the address.
The term postcode, rather than postal area, is used in the text, for ease of reading.

Postcode
See postal area, above.
Publish with Bio Med Central and every scientist can read your work free of charge

SLA
An SLA in Perth is generally equivalent to a local government area, with additional codes allocated to local government areas split for statistical purposes (mainly local government areas with large populations, split to form SLAs with smaller populations).