Hierarchical Models for Multiple Outcomes

Abstract

Hierarchical Models for Multiple OutcomesAbstract Number:2496 David Richardson* David Richardson* UNC SPH, United States, E-mail Address: [email protected] Search for more papers by this author AbstractIn an occupational or environmental cohort study, there may be interest in associations between an exposure primary interest and mortality due to a variety of causes. A standard approach involves fitting a separate regression model for each outcome type. However, estimated associations may be imprecise if there are few events for each outcome type. A commonly-used strategy to deal with sparse outcomes is to combine several outcome types into a broader category, such as all solid cancers, and perform regression analysis on this broader outcome group. However, this strategy does not allow inferences regarding effects of exposure on specific outcome types, is sensitive to decisions about how to combine outcome types, and imposes the assumption of homogeneity of the exposure effect across the combined outcome types. We describe an alternative to coalescing several outcomes: a hierarchical model for cause-specific hazards that can be used to estimate an exposure’s effects on different causes of death, while stabilizing the parameter estimates of interest. We address hierarchical models for estimation of the main effects of exposure, as well as for estimation of parameters that describe effect measure modification. The approach is illustrated in analyses of solid cancer mortality in a cohort of dioxin-exposed chemical workers. Compared to standard estimates of cause- specific exposure-mortality associations, hierarchical regression yielded estimates with improved precision that tended to have less extreme values. Moreover, the hierarchical approach allowed estimation of models with parameters to describe cause-specific effect measure modification, while estimates of these parameters were not possible to obtain in standard models fitted one outcome at-a-time. The proposed approach can yield estimates of association that outperform conventional regression methods when one wishes to estimate associations with multiple outcome types.