Salmerón et al. Respond to “Future Directions for Predicting Rare Cancer Rates”

We thank Buller and Jones (1) for their careful reading of our paper (2). They suggest potential strengths and weaknesses of our Bayesian modeling with regard to its future extension to a spatial/temporal framework. As they noted, identifying causes of and developing strategies for prevention of rare cancers (RCs) is a unique challenge (1). It requires reliable estimation of the burden of these cancers across areas with cancer registry data, and this confronts the challenge of variations in population coverage and data quality in each specific cancer surveillance system (3–5). In our paper (2), we presented statistical modeling for obtaining RC incidence estimates from national/regional public health organizations for approximately 190 cancer entities in about 20 European countries, not all of them with national registry coverage (3–5). The lack of information on RCs affects public health planning (4). Since such expertise is crucial for clinical management of RC (3, 4), rational centralization of cases is a critical issue, especially for small and medium-sized countries.

In their commentary (1), Buller and Jones pointed out a double issue with regard to RC statistics: statistical uncertainty and potential identifiability of patients. The Bayesian approach could tackle both problems by incorporating, at best, the information available on the incidence rate distribution and thus providing (unidentifiable) estimates instead of (identifiable) raw numbers. Since focusing only on the estimation methods of RCs might present complex technical challenges (1, 2), there must be an epidemiologic perspective with which to build a set of statistical indicators taking into account potential underestimation of incidence rates (2) and their spatial variability (1). As Buller and Jones stated (1), the at-risk population and risk factors could be spatially heterogeneous for each RC, and thus estimation of the burden of each RC might require a specific a priori setting (6–8). Collection of this information in future studies is of particular interest.

The model we presented in our study (2) can be extended for both within- and between-country variability and to include risk factors and time trends as covariates. The model could also be used for predicting RC incidence (1) in areas without cancer registration by using information on the RC incidence and mortality in nearby areas (9). However, the use of structured variability, which makes epidemiologic sense when applied to small adjacent areas within the same country, is harder to justify when comparing countries (5). Appropriate assumptions and considerations must be made when coupling the numerical analysis of the prior distributions with the exploration of etiological hypotheses, presenting a challenge in the statistical modeling of RC which should be closer to epidemiologic thinking.

We addressed overdispersion in our Poisson model by studying different lower bounds of a uniform hyperprior distribution for the precision of the random effects (2), showing that the performance of this model covers the scope of scenarios observed in our data (2, 5). However, as Buller and Jones noted (1), future studies must explore probability distributions other than the Poisson distribution in order to account for unconsidered correlation in the data used for estimating the burden of RC. These must take into account spatial or temporal variability (or both) and excessive zero counts (7, 10) in the data. This last situation is common when estimating the incidence of RC, even in areas with large population (5). Therefore, more studies comparing the performance of these probability distributions are called for.

As Buller and Jones concluded (1), there is still much work to be done in identifying strategies for tackling the challenge of estimating the burden of RC. The provision of indicators based on flexible modeling is a step in this direction.

ACKNOWLEDGMENTS

Author affiliations: Health and Social Sciences Department, Instituto Murciano de Investigación Biosanitaria Virgen de la Arrixaca, University of Murcia, Murcia, Spain (Diego Salmerón); Centro de Investigación Biomédica en Red Epidemiología y Salud Pública (CIBERESP), Murcia, Spain (Diego Salmerón); Evaluative Epidemiology Unit, Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy (Laura Botta, Annalisa Trama, Gemma Gatta, Riccardo Capocaccia); Department of Statistics and Operations Research, Technical University of Catalonia, Barcelona, Spain (José Miguel Martínez); Public Health Research Group, University of Alicante, Alicante, Spain (José Miguel Martínez); Cancer Plan, Institut d’Investigació Biomèdica de Bellvitge, Catalan Institute of Oncology, Hospitalet de Llobregat, Spain (Josep M. Borràs, Ramon Clèries); and Clinical Sciences Department, Faculty of Medicine and Health Sciences, University of Barcelona, Barcelona, Spain (Josep M. Borràs, Ramon Clèries).

This work was supported by the European Commission through the Consumers, Health, Agriculture and Food Executive Agency (grant 2000111201). We acknowledge the support of the Agència d’Avaluació d’Universitats i Recerca (grant 2017SGR00735) and the Centres de Recerca de Catalunya (CERCA) Program of the Generalitat de Catalunya. We also acknowledge support from the Fundación SéNeCa—Agencia de Ciencia y Tecnología de la Región de Murcia Program for Excellence in Scientific Research (project 20862/PI/18).

The funders played no role in the study design; data collection, analysis, and interpretation; or the writing of the report.

Conflict of interest: none declared.

REFERENCES