We present two novel models for estimating rates of sensitive behaviors among dyadic units in the presence of disagreement about the occurrence of these behaviors within the dyad. One model treats true status as known for concordant dyads and estimates the associations of the behavior with covariates. These relationships are then extrapolated to the discordant dyads to improve the baseline incidence estimate for this group, which must be supplied externally in the absence of a gold standard. We adopt a Bayesian approach by specifying this baseline estimate in the form of a prior distribution on missing data. Alternatively, we can model individual reports as a function of latent status (treated as unknown for all dyads) and a latent propensity to report. Either latent variable may depend on covariates. Comparisons of these models to each other and to other available methods are presented. The models are illustrated with violence, substance abuse, and demographic data from a large population-based survey of cohabiting heterosexual couples.
Learning Objectives: Participants will be introduced to new methods for estimating rates in the presence of possibly discordant reports from multiple informants. Following this presentation, participants should be able to identify the strengths and weaknesses of these new methods relative to existing techniques in terms of strength of various assumptions, flexibility in handling missing data, inclusion of covariates, and computational complexity.
Keywords: Risk Behavior, Statistics
Presenting author's disclosure statement:
Organization/institution whose products or services will be discussed: None
Disclosure not received
Relationship: Not Received.