DOSE-RESPONSE MODELING WITH MARGINAL INFORMATION ON A MISSING CATEGORICAL COVARIATE

When the relationship between a dosage-type variable and a binary outcome depends on a categorical variable, a common analysis would employ a dose-response model with the categorical variable as a covariate. When the level of the categorical variable is not known for all subjects, however, the standard dose-response model alone cannot provide useful inference. We present an EM-based approach to account for the missing covariate in a dose-response model setting when additional knowledge about the marginal distribution of the covariate is available. This approach is motivated by a study of the beetle Rhyzopertha dominica, a pest of stored grain in Australia. Certain genotypes of this beetle have developed inheritable resistance to a widely-used insecticidal fumigant. In this study, the effects of various dosage levels of the fumigant were considered, and it was feasible to genotype only the surviving beetles.