Far East Journal of Theoretical Statistics
Volume 16, Issue 2, Pages 285 - 302
(July 2005)
|
|
USE OF AN ARTIFICIAL NEURAL NETWORK IN EXPLORING THE DOSE-RESPONSE RELATIONSHIP BETWEEN CIGARETTE SMOKING AND LUNG CANCER RISK IN MALES
Karin Bammann (Germany), Hermann Pohlabeln (Germany), Iris Pigeot (Germany) and Karl-Heinz Jöckel (Germany)
|
Abstract: The
dose-response relationship between smoking and
lung cancer risk is typically modelled using
generalised linear models, thus assuming a
certain structure of this relationship. In
logistic regression models, these assumptions
imply a linear effect of a risk factor on the
log odds ratios and independent effects of
different risk factors on the disease risk.
One possibility to meet violations of model
assumptions is appropriate transformations of
the data before modelling, another possibility
is the employment of a more flexible
structural model. An artificial neural network
was used to explore possibly neglected aspects
in the dose-response relationship between
smoking and lung cancer risk. The results were
compared with those obtained by analysing the
data with classical logistic regression
models. The results show a non-symmetrical
effect of dose and duration of smoking on the
lung cancer risk, identifying dose as the main
risk factor for lung cancer in males with low
cumulative smoking exposures and duration for
highly exposed males. The results strongly
encourage the use of artificial neural
networks in analysing epidemiologic
case-control studies, in spite of the complex
task of interpreting the results and the
lacking statistical safety of the models. The
benefit to gain is an essentially more
differentiated picture of dose-response
relationships than classical regression models
can supply. |
Keywords and phrases: case-control studies, cigarette smoking, dose-response relationship, lung cancer, neural network models. |
|
Number of Downloads: 263 | Number of Views: 748 |
|