An approach to estimating exposure-specific rates of breast cancer from a two-stage case-control study within a cohort

The Breast Cancer Detection and Demonstration Project (BCDDP) included a large cohort of women followed for incidence of breast cancer and from whom an initial case-control sample was drawn and standard risk factors obtained. In order to study the effect of mammographic features on breast cancer risk, a nested subsample of cases and controls was drawn. Therefore, these data can be viewed as two-stage case-control data within a cohort, or as cohort data with two nested levels of missingness, since basic characteristics like age were measured on all members of the cohort, standard risk factors were elicited only in the initial case-control sample, and mammographic features were assessed only in the nested subsample of cases and controls. We present a Poisson pseudo-likelihood approach to estimating age- and exposure-specific breast cancer incidence rates based on the three types of variables. This approach takes into account the nested missingness as well as two other type of missingness, namely, that for basic variables and standard risk factors, some levels (i) were omitted by design in the nested subsample of case and controls or (ii) were empty because of the sparsity of the data in that subsample. Estimates of standard errors are obtained from a parametric bootstrap. The approach seems to be efficient when applied to the BCDDP data and is flexible for modelling breast cancer rates and taking the special missingness features of these data into account.