Flexible statistical models provided new insights into the role of quantitative prognostic factors for mortality in gastric cancer

Objectives: To reassess the effects of prognostic factors on mortality in gastric cancers, and to illustrate the advantages of flexible modeling. Study Design and Setting: A prospective population-based cohort of persons diagnosed with gastric cancers in 1976 to 1995 in Burgundy, France, was followed for 5 years since diagnosis. Multivariable survival analyses, stratified by cancer stage, involved both conventional Cox's model and its flexible generalization, which permitted testing the underlying assumptions and accounting for changes over time in the effects of prognostic factors. Results: Conventional assumptions of proportional hazards (PH) (P = 0.003) and linear increase in risk with increasing age (P = 0.003) were rejected for age at cancer diagnosis in patients with less advanced (stage I/II) gastric cancers. Flexible modeling revealed that: (1) the effect of age was important mostly in the first year after diagnosis; and (2) mortality increased both for the youngest and the oldest patients. In conventional analyses, calendar year of diagnosis had no association with mortality in stage I/II cancers (P = 0.37). In contrast, flexible analyses indicated a statistically significant reduction in early, mostly postsurgical, mortality between early 1970s and 1980s. Conclusion: Flexible modeling and testing conventional assumptions may yield new insights in prognostic studies of cancer mortality. (C) 2008 Elsevier Inc..All rights reserved.