Multilevel Modeling of Hormonal Factors and Breast Cancer
John Witte
Department of Epidemiology and Biostatistics, CWRU
In contrast with a conventional analysis, multilevel modeling can give more
stable and plausible estimates by incorporating additional information in a
higher-level model. This approach also offers a solution to problems of
multiple inference. Here, we show how multilevel modeling can be used to
estimate the relation between breast cancer and "hormonal factors" (i.e.,
age at first full term pregnancy, oral contraceptive (OC) use, body mass
index, parity, and years menstruating). We use data from a case-control
study, where cases are women with premenopausal bilateral breast cancer
(Ncases = 140) and controls are their non-diseased sisters (Ncontrols =
222). A conventional analysis of these data entails conditional logistic
regression of breast cancer on the hormonal factors.
Using this approach, for example, the odds ratio (OR) comparing OC users
versus non-users equals 1.7 (95% Confidence Interval (CI) 1.0-2.8). This
analysis, however, ignores the similarities among these factors with regard
to their potential effect on breast cancer. In particular, the hormonal
factors' effects on breast cancer may each have components due solely to
the proliferation of breast cells, which parallel the rate of breast tissue
aging (Pike et al., Nature 1983;303:767-770). Therefore, in an attempt to
improve the conventional estimates, we use a multilevel model in which the
hormonal factors' effects depend linearly on their relation to the
proliferation of breast cells. Using this model, the OR for OC use now
equals 1.3 (95% CI 0.9-1.9). This work demonstrates how multilevel models
can be developed for use in epidemiology, and how this approach can result
in more precise and plausible estimates than a conventional analysis.
Questions? Jiming Jiang