Abstract : Abstract
Survival data analysis and reliability have many common points and the concepts that are studied are comparable. But the data have often different properties as we are not in the same experimental field. For example, we cannot do "accelerated" experiments on patients. Also the patients are not only "objects" of study but also "subjects". And while we try to compare the effects of two drugs to cure a particular disease, the patient may decide at any time to stop his participation in the experiment, which leads to what we call "right censoring" of the data which are only partially observed. If we add to this feature the possibility of "truncation" (only a part of the sample is observed, some subjects are not included in the sample) and the fact that the size of the patient sample is often limited to small numbers, we can see that there are multiple differences in the available data in survival analysis as compared to reliability analysis.
However, the same models can be used. For example, even if we cannot "accelerate" experiments on patients by enforcing the "stresses" applied to them, we can use such an accelerated model in order to understand whether a specific exposure to pollution is accelerating the onset of a specific disease. We give here an example where the occupational exposure to asbestos is shown to accelerate the onset of lung cancer.
The aim of this talk is to show how one can deal with the censoring and truncation inconveniences, in both case of only one terminal event(1,3,4) and also when several different outcomes may happen(5,6). But, as the aim of survival analysis is not only to try to increase the lenght of the pure survival time (to death) but also the survival time "free of disease", this leads to estimate the number of "years free of disease" lost(2,7) due to environmental or behavioural factors for the sake of prevention.