Cancer Epidemiol Biomarkers Prev. 2021 Jun 22:cebp.1727.2021. doi: 10.1158/1055-9965.EPI-20-1727. Online ahead of print.
ABSTRACT
BACKGROUND: The American Cancer Society (ACS) and the National Cancer Institute collaborate every 5-8 years to update the methods for estimating numbers of new cancer cases and deaths in the current year in the U.S. and in every state and the District of Columbia. In this paper, we re-evaluate the statistical method for estimating unavailable historical incident cases which are needed for projecting the current year counts.
METHODS: We compared the current county-level model developed in 2012 (M0) to three new models, including a state-level mixed effects model (M1) and two state-level hierarchical Bayes models with varying random effects (M2 and M3). We used 1996-2014 incidence data for 16 sex-specific cancer sites to fit the models. An average absolute relative deviation (AARD) comparing the observed to the model-specific predicted counts was calculated for each site. Models were also cross validated for 6 selected sex-specific cancer sites.
RESULTS: For the cross-validation, the AARD ranged from 2.8% to 33.0% for M0, 3.3% to 31.1% for M1, 6.6% to 30.5% for M2, and 10.4% to 393.2% for M3. M1 encountered the least technical issues in terms of model convergence and running time.
CONCLUSIONS: The state-level mixed effect model (M1) was overall superior in accuracy and computational efficiency and will be the new model for the ACS current year projection project.
IMPACT: In addition to predicting the unavailable state-level historical incidence counts for cancer surveillance, the updated algorithms have broad applicability for disease mapping and other activities of public health planning, advocacy and research.
PMID:34162657 | DOI:10.1158/1055-9965.EPI-20-1727