Using modern Bayesian small area modelling and mapping techniques we have been able to show that all investigated groups of women in Switzerland have benefited from progress in cancer control regardless of place of residence in the past 40 years. We observed only small differences in the geographical variation of mortality.
A factor, which may have contributed to breast and uterine cancer mortality reductions, is the change in the use of hormone replacement therapy (HRT) [7]. After an association of HRT use with breast cancer occurrence was reported [8], its use declined sharply.
We were also not able to show similar spatial patterns in breast and ovarian cancer mortality although they share several life style related, environmental and genetic risk factors. It should be noted however, that hereditary cancer accounts only for about 5-10 % of the cases in breast cancer [9] and about 15 % in ovarian cancer [10]. They are shown to occur at younger age and more advanced stage; still, a visible effect on the mortality map may only be seen in areas with ethnic groups or very large families with a highly elevated risk for hereditary cancer. Such a risk has been described for Ashkenazi Jewish women. The BRCA Ashkenazi founder gene mutations are prevalent in approximately 2 % of these women [11] with communities of Ashkenazi mainly found in urban areas; largest communities are in the cities of Zürich, Geneva and Basel contributing to 1-2 % of the population [12, 13]. However, the breast and ovarian cancer risk in BRCA carriers is affected by genetic modifiers and non-genetic factors, for example, reproductive behaviour, hormonal exposure, lifestyle and risk reduction surgeries [14]. We could not observe an elevated mortality for the three cities in contrast to the surrounding area and it remains unclear to which extent the mortality rates are driven by these hereditary forms of cancer.
Considerable differences in health and health related behaviour have been reported for the Swiss language regions including alcohol intake, smoking and a healthy diet [15, 16] but lacked significance as regression factors in our analysis.
Only for three cancer site-age group combinations was the urbanisation level identified as a significant factor. Urbanisation is serving as a proxy for access to and quality of medical services, education and health consciousness [3]. By our regression with 20 years of new data, we could not formally confirm an urban–rural gradient for breast cancer as described by Schüler & Bopp [3] as significant.
Overall, no general pattern across age groups or cancer sites was present.
The reduction of mortality was stronger in the younger age groups, which is probably the result of better survival and therefore a shift in the age of death. This would also explain the temporary increase in breast and ovarian cancer death risk around the year 1990 in the 75+ year olds. In addition, in this age group multi-morbid conditions and fewer treatments are common [17]. Sant et al. [18] noted that poor survival for gynaecological cancers in the elderly could be due to advanced stage at diagnosis, or failure to give adequate treatment, perhaps because of comorbidity. In general, the interpretability of results in this age group is limited due to its small size, more multi-morbid conditions together with possible inconsistencies in death certification over time, because of only allowing one single cause of death.
Strengths and limitations
As cancer deaths are rare events and in order to increase the power, different geographical units have been used when analysing cancer mortality data in the past. Some authors have used selected cantons [2] and Schüler & Bopp [3] used for their cancer atlas somewhat smaller mobility regions based on the accessibility to goods and services but which do not take into account population size. As a result, this choice was too aggregated for some urban areas and not aggregated enough for some sparsely populated areas in order to reveal robust, underlying trends. In view that the choice of the geographical unit of analysis may greatly influence results [19], the combination of small geographical units with a state-of-the art smoothing technique enabled a more detailed analysis. With this analysis, we could additionally show the driving age groups or subareas of elevated or reduced mortality in certain regions, while reducing uncertainties due to small numbers and adding an investigation of non-linear time trends.
In general, smoothing allows an estimation of the underlying risk, in a sort of a long-year average, rather than the actual situation. However, for single municipalities, without fully eliminating it, the use of Bayesian smoothing reduces the probability to detect narrow areas with specifically high or low risk. Municipalities at the country border may not benefit from smoothing to the same extent as municipalities in the interior of the country due to unknown data on the other side of the border. Therefore, in the interpretation of the results emphasis should be given to the broader spatial patterns rather than to single municipalities.
Comparing with the previous work of Schüler & Bopp [3] our study not only extended their work by 20 more years and corrected for non-linear time effects, more importantly, we were able to correct the foreseen overestimation in mortality numbers until 1994, which could not be adequately addressed earlier. Priority rules in the coding of causes of death led to an overestimation in cancer deaths due to their prioritization over other comorbidities. The applied methodology of age standardisation takes advantage of the actual age structure rather than a standard population.
There are important limitations to our study. Risk factors affect incidence but are not necessarily linked to mortality [20]. The progression stage of the tumours and their histological type could not be taken into account, as the ICD-classification does not include histological type for the sites studied. The regional case mix and its changes over time therefore may have distorted the results.
Further distortions may arise from the uncertainty as to what level the reported main cause of death and comorbidities are comparable in time and between regions, although the central coding speaks in favour of a certain homogeneity in the coding procedure. In the elderly with frequent multi-morbid conditions, the probability of misclassification is higher.
Furthermore, after prior analysis the covariates language region and urbanisation level were fixed in time for the municipalities, so that varying developments therein may have resulted in inaccuracies.