Part II: Inequalities in oral cancer in relation to relative deprivation of area
4.5.1 Incidence in relation to deprivation 1 Data processing method
TCR registration data were aggregated to the level o f the 1991 Census Enumeration District (ED) and combined with the Townsend index as described in Chapter 3. Population denominators were derived from the 1991 Census. Most patients’ records included details o f their home postcode at the time o f diagnosis, and the All Fields Postcode Directory (AFPD) was used to assign patients to their respective EDs. Within each ED, numbers o f patients were stratified by gender, age group, and tumour site.
Enumeration districts were chosen as the geographical units for the analysis as they are the smallest units available from the Census. People living within a given ED will be more homogeneous and more likely to share similar living conditions than will people living in larger geographical areas. EDs therefore provide the finest level o f resolution for the study o f any associations with relative deprivation.
The Townsend index was chosen as it is widely used in the study o f relative deprivation and inequalities in health. It has been extensively validated (Morris and Carstairs, 1991).
4.5.1.2 Data processing - method 2
For the second method, the data were aggregated to the level o f the 1998 Ward. The population denominators were the mid-1998 population estimates provided by Oxford University as part o f the project that developed the Index o f Multiple Deprivation 2000 (IMD 2000) for the Department o f Environment, Transport and Regions (DETR). Deprivation was classified according to the IMD 2000. The
population estimates were not provided for males and females separately. It was therefore not possible to stratify the analyses by gender. Only three broad age groupings were available for age standardisation (0-15 years old, 16-59 years old, 60 years old and above).
The Ward-based IMD 2000 and the corresponding population estimates were chosen to see whether this approach was valid when compared to the use of the Townsend index recorded at the ED level. This was necessary since data for later analyses in this project (dental registration data, and Hospital Episode Statistics) were not available at the level o f the ED and covered a slightly later time frame.
4.5.1.3 Statistical analyses
For both methods described above, the STATA statistical software package (STATA Corporation, 2002) was used to undertake Poisson regression analyses in order to calculate the age-adjusted Incidence Rate Ratios (IRR) o f oral cancer at differing sites in relation to area deprivation. In each case the area deprivation characteristics were entered into the models as quintiles o f the relevant index (either Townsend or IMD 2000) with the least deprived quintile being the baseline group for all comparisons. All models included the population estimate as an offset to ensure that differing population sizes were controlled for in the analyses (Breslow and Day, 1993a).
The effect o f quintile o f deprivation was considered as two nested models for each site combination: firstly as four dummy variables, and secondly as a continuous scale. For each set o f nested models the likelihood ratio test was used to determine whether the dummy variable approach explained significantly more o f the variation
in cancer incidence than if the effect o f deprivation was assumed to follow a linear relationship. This was to see whether a simple model (in which the steps between quintiles were treated as being equal) was an adequate method o f describing the relationship between cancer incidence and relative deprivation, or whether the relationship was more complex.
Using method 1 (Table 4.6) it was possible to produce results for males and females separately. Method 2 (Table 4.6) only permitted results to be calculated for both sexes combined. In order to determine the extent to which the IRR estimates were similar from both approaches (and therefore validate the IMD 2000), a further set of analyses were undertaken. These used the data from method 1, but IRRs were calculated that were simultaneously adjusted for age and sex, instead o f calculating age-adjusted estimates for males and females separately.
4 .5 .2 Tumour stage in relation to deprivation 4.5.2.1 Data processing - method 3
As before, the AFPD was used to assign patients to their ED of residence in order to assign each patient a Townsend index score. However, the records were left as individual patient data rather than being aggregated. Quintiles for the Townsend score were calculated based on this data set. In methods 1 and 2 (Table 4.6) the quintiles were based on dividing the total number o f geographical areas (EDs or 1998 Wards respectively) into five approximately equal groups based on the index score. In methods 3 and 4, the quintiles were produced by dividing the total number o f individual cancer patients into five approximately equal groups based on the index score.
4.5 2.2 Data processing - method 4
The data processing undertaken was identical to method 3 except that individual patients’ deprivation quintiles were based on their Ward IMD 2000 score.
4.S.2.3 Statistical analyses
A variable was generated to indicate whether patients had early or late stage disease at the time o f presentation. Patients classified by TCR as stage 1 or 2 were considered to have ‘early’ disease. Those with TCR stage 3 or 4 were considered to have ‘late’ disease. Patients for whom stage information was not available were excluded fi"om the analyses. Logistic regression was utilised to investigate the possible relationship between the age-adjusted odds o f late presentation and quintile of deprivation separately in males and females. Thus the exposure was deprivation quintile and the outcome o f interest was ‘late’ disease with least deprived quintile being the baseline group for comparisons (Breslow and Day, 1993b). This was principally to see whether there was any evidence that people living in more deprived circumstances presented later in the disease’s natural history.
As in the aggregated analyses, the regression was undertaken alternately assuming and not-assuming a linear relationship between the outcome (odds o f presenting late) and the quintile o f deprivation. Likelihood ratio tests were used to compare the fit of nested models. This was to determine whether or not a simple model, in which the difference between deprivation quintiles was assumed to be equal, was an adequate description o f the relationship between late presentation and relative deprivation.
Some cases had to be excluded from the analyses due to lack o f availability o f stage information. The data were checked to ensure that there was no differential in the
reporting o f stage information across deprivation quintile. If such a differential had existed it could have biased the results. In order to check this a chi-squared test was used to investigate whether there was any evidence o f a difference in reporting of tumour stage between quintiles o f Townsend deprivation score. A Wilcoxon rank sum test was used to determine if there was a systematic trend in differential reporting across the quintiles.
4.6 Results
4.6.1 Incidence in relation to deprivation