APH 2002, 60, 73-100:
Description of the influence of age, period and cohort effects on cervical cancer mortality by loglinear Poisson models (Belgium, 1955-94).
M. Arbyn, H. Van Oyen, F. Sartor, F. Tibaldi and G. Molenberghs
Keywords: age-period-cohort models, trend analysis, poisson regression models, mortality, cervical cancer, Belgium
Background: Cervical cancer mortality in Belgium has been decreasing continuously over the last forty years. This might generate the impression that the trend has hardly been influenced by changing exposure to etiologic factors or by increasing attendance to screening conducted since twenty years. It is important to separate out the role of ageing, period of death and period of birth (cohort).
Method: An age-period-cohort analysis, based on Poisson regression, was performed on cervical cancer mortality in Belgium between 1955 and 1994 in women between 20 and 79 years. The method of model building as proposed by Clayton (1,2) is used. A linear secular trend (drift) can be isolated but not attributed to either period- or cohort-effects. Only the non-linear deviations are estimable using second differences contrasts. Overdispersion is allowed.
Results: The mortality decreased with about 50% over the last four decades. A full age-period-cohort model, adjusted for extra-Poisson variation, was necessary to adequality describe the trends. Strong cohorteffects were observed, besides age and drift. The non-linear period effect was significant but limited in magnitude.
Conclusions: The cohort effects seem to coincide with changing sexual behaviour of successive generations. The existence of a substantial negative drift factor shows that the decrease of mortality cannot be ascribed simply to prevention by Papanicolaou testing. Otherwise it does not provide evidence that screening was not influential. It is possible that screening further prolonged the effect of earlier clinical diagnosis and treatment due to improved access to health care.