The accessibility of a person's residence was the greatest predictor of an increased risk of cancer diagnosis across a range of cancers, including lung (females), melanoma, breast (females), cervical, prostate, and non-Hodgkin's lymphoma. Socioeconomic status was the greatest primary explanatory variable for lung cancer (males).
More remote areas had a greater probability of having high incidence of lung cancer among males, and cervical cancer. Cancers for which more urban areas were more likely to have high incidence included: lung cancer (females), melanoma, breast cancer, prostate cancer, and non-Hodgkin's lymphoma.
The interaction between accessibility, socioeconomic status and ethnicity varied depending on the type of cancer. The socioeconomic status interacted with accessibility for lung, melanoma, breast (females), cervical, and prostate cancers. The incidence of cancers that were often screen detected such as breast cancer (females), melanoma (males) and to a lesser extent prostate cancer tended to be higher in more affluent areas, and also more urban areas. In contrast, for lung, melanoma (females) and cervical cancer the incidence was higher in more disadvantaged areas. Cancers with a high incidence in disadvantaged areas did not have a consistent interaction with accessibility. Some tended to be higher in more urban areas (such as lung cancer (females) and melanoma (females)), while others were higher in more remote areas (lung cancer (males) and cervical cancer). Ethnicity also interacted with these factors for cervical cancer, with Indigenous areas more likely to have high incidence.
These results are consistent with previous studies showing an increased incidence of cervical cancers among Indigenous women [21], and an increased incidence of breast cancer among women in more urban or affluent areas [22]. However, there are also important differences compared to previous research. Melanoma incidence has generally been found to be higher in more affluent areas [23]. In contrast, our results found females in the most advantaged areas were less likely to have high incidence, while all other SLAs (except for very remote) were more likely to have high incidence. Queensland has among the highest rates of melanoma in the world [3, 24], and this may be impacting on these differences. Similarly, lung cancer incidence has previously been shown to be higher in remote areas for both males and females [25]. However, our results found high incidence among females in the lower socioeconomic areas of major cities.
Individual risk factors could be influencing these geographic differentials. Lung cancer incidence is strongly determined by smoking prevalence 20-30 years earlier [26]. Tobacco smoking has been shown to be more prevalent in lower SES or more remote areas, which may explain the high incidence observed in these areas [27–32]. Similarly, women in affluent areas are more likely to delay childbearing, have fewer children and/or use hormone replacement therapy, all of which are risk factors for breast cancer [33–35].
Preventive measures can also differ geographically. The leading cause of cervical cancer is infection with sexually transmitted human papillomaviruses. Papanicolaou screening (commonly called pap smear testing) detects precancerous lesions, which can then be treated, averting cancer and thus lowering incidence. The high incidence observed in very remote, Indigenous or the most disadvantaged urban areas may result from lower uptake of pap smears. Participation rates for cervical cancer screening (papanicolaou screening) are lower in remote communities and areas of low socioeconomic status in Queensland and throughout Australia [36, 37].
In contrast, screening for asymptomatic cancers, such as prostate or breast cancer, can be associated with increased incidence. Therefore access to screening or diagnostic services is another factor which influences incidence and can vary by area. For instance, the incidence of prostate cancer may be inflated in areas where prostate-specific antigen (PSA) testing, which is used to detect asymptomatic prostate cancer, is commonly used. PSA testing is less common in more rural areas than in capital cities throughout Australia [38], and this could be contributing to the lower incidence in remote areas. Breast cancer may also be influenced by geographic variation in screening services, as there is variation in mammogram uptake by accessibility and socioeconomic status [39]. Similarly, the ease of access to skin cancer checking services in more urban areas may influence the incidence of melanoma.
Strengths of the study include the use of routinely collected incidence data from a population-based registry to which notification of cancer is required by law. Queensland has the most decentralized population in Australia [40], thus providing a unique opportunity to investigate these area-based differences in greater detail.
Limitations of the study include the nature of cancer, which takes years to develop and be diagnosed. Therefore it is possible that the incidence of an area may reflect the risk factor prevalence from years earlier, rather than the current situation. Also, estimates were calculated based on area of residence at diagnosis. People may have migrated to different areas leading up to their cancer diagnosis, and any carcinogenic exposure or other area-level influences may have occurred at a different location to where they were diagnosed.
The CART analysis was weighted by the inverse of the variance, which had the effect of placing greater priority on correctly identifying SLAs with high SIRs (or sensitivity), so the specificity (correct identification of SLAs with non-high SIRs) was found to vary considerably between cancers and gender. Two cancers with comparatively low sensitivity and specificity were prostate cancer and male melanoma. Therefore, results for these models should be treated with caution.
The 'high' SIR values were classified as an arbitrary cut-off of at least 10% above the Queensland average. This value was chosen to increase the probability that results were truly above the State average values. Since it was probable that choosing alternate cut-off values would influence the tree structure, sensitivity analyses (not shown) were performed under alternate cut-offs (5% and 15% above the Queensland average). Although different cut-off values often induced some variation in tree structure, the primary split remained identical for all cancers except for minor differences in the categories included on either side of the split for male lung cancer, female breast cancer, cervical cancer, prostate cancer and male non-Hodgkin's lymphoma.
Since the incidence of some cancers such as breast, melanoma and prostate is strongly influenced by screening practices, high incidence may result from overdiagnosis, where asymptomatic cancers are detected which would not otherwise have progressed to cause morbidity and/or death. While in this case a high incidence of cancers may not necessarily be an adverse outcome in itself, the morbidity associated with subsequent treatment is sometimes considerable [41]. Similarly, low incidence may not necessarily be beneficial if the cancers which are diagnosed are detected at a more advanced stage and therefore have worse prognosis. Consistent with other Australian Cancer Registries, the QCR does not routinely collect staging information for all cancers. Therefore it was not possible to differentiate between areas at high risk of having advanced cancers diagnosed, and those at high risk of having sub-clinical cancers diagnosed.
Alternative methods are available to explore interactions. For instance, increasingly cancers are jointly modelled, either using multivariate structures on the relative risks, or latent class models [42]. One benefit of these methods is utilizing strength between the cancers to produce more efficient estimates [43]. By exploring spatial variation in common risk factors, latent class models can provide stronger evidence of any true clustering in the underlying risk surface [43]. However, under latent class joint modeling the shared components (risk factors) for each cancer are pre-specified, whereas the CART analysis determines which of the risk factors are relevant for that cancer. The use of different modelling strategies may identify different features of the data that can lead to better understanding of the problem at hand and can thus lead to more informed inference. For example, in addition to being a valid approach in its own right, a CART model may identify useful interactions for inclusion in a subsequent (univariate or multivariate) regression analysis.
where p is the Pr(H) and n is the number of SLAs. In the few instances where a CI value surpassed the possible (0,1) boundaries, this was restricted to the appropriate boundary value. The CART diagrams are also accompanied by summary diagrams showing which areas were likely to have high SIR values (shaded as dark grey), and which were likely to not have high SIR values (shaded as light grey). These contain ARIA and SEIFA combinations to facilitate comparison between cancer types. Combinations which do not exist were rendered in white. Note the same shading is also used for the terminal nodes in the CART diagram. Dark grey terminal nodes are likely to have a high SIR, in contrast to the light grey terminal nodes.
