gms | German Medical Science

The occurrence of malignant neoplasms is observed by cancer registries in many regions all over the world. Hence, the comparison of cancer incidence between different regions is the objective of many studies. Yet, there are some methodical issues that need to be addressed in such analyses.

Since the probability of a cancer diagnosis depends on age, the distribution of which varies between different regions, standardization for age is required. There are two standardization approaches. In direct age standardization, local age-specific rates are applied to a standard population. The resulting age-standardized rates (ASR) may be compared between the regions. In indirect age standardization, age-specific rates of a reference population are applied to the local population structure, providing the expected number of cases. Then, the ratio of the observed case number to the expected number gives the standardized incidence ratio (SIR). The choice between ASR and SIR depends on the data at hand. Usually, the ASR is preferable, but the SIR is favourable for small sample sizes.

Further challenges arise from the conflict between the size of the regions under study and the accuracy of the corresponding incidence estimates. On the one hand, the aggregation into few large regions may hide within-regional variations. On the other hand, small sample sizes resulting from small study regions cause large variances and therefore a “patchy” map. A solution for this conflict is the application of spatial smoothing methods. Unfortunately, only smoothing methods for SIRs are available in cancer epidemiology by now.

In our research, we modified the algorithm of Besag, York and Mollié to allow its application to directly standardized rates. In its original definition, this model assumes the observed cases in each region to follow a Poisson distribution. The parameter of the Poisson distribution (lambda) is the expected case number estimated from the overall population, multiplied by a risk factor consisting of an individual and a spatially auto-correlated term. In our approach, we adopted the latent risk factor but suggest to assume a log-normal distribution for the ASRs instead of the Poisson distribution used for the SIRs.

In this contribution, we would like to introduce briefly the method of age standardization and the concept of spatial smoothing. Following, our approach will be described and demonstrated by an application to real-world cancer incidence data.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.