Multinational trends in cancer mortality rates. Methodological issues and results.

Identification of trends in cancer mortality rates for specific sites is useful both for targeting areas of health concern and for suggesting topics where epidemiologic research into etiology might be fruitful. Differences and similarities across countries may provide clues for further research. Counts of cancer mortality by sites are often considered as Poisson processes. Using this framework, Poisson regressions can be used to identify time trends. Model identification can also be difficult and is aided by the graphical techniques of exploratory data analysis. Plots of the data can identify the existence of trends, nonlinearities, cohorts with special risks, and artifacts due to such things as ICD revision. Because human perception is often distorted by outliers in the data, local nonparametric smoothing can yield more robust plots for model identification. Quasi-likelihood methods can be used to incorporate serially correlated and overdispersed covariance structures. These techniques are illustrated using cancer mortality data from selected industrialized countries.