Abstract:
|
Environmental exposure effects on human development cannot be studied by experiments, but may be small and difficult to detect in observational data. In the Seychelles Child Development Study (SCDS), researchers examined the effect of prenatal methylmercury exposure on a battery of 20 tests measuring aspects of child development (Thurston et al., 2009; Xiao et al., 2014). Thurston et al. posed a nested model where covariate and exposure effects differ across domains and two levels of random effects capture correlations in the same domain and across all outcomes. Xiao et al. extended the model by allowing mixed-membership in domains. We use model-based clustering to group the outcomes into an unknown number of domains with a Dirichlet Process prior. Estimation is a Bayesian MCMC algorithm combining univariate and split-merge steps. Convergence is checked using multivariate effective sample size. Simulations show good power to determine the number of domains. Results from the SCDS show more pronounced methylmercury exposure effects in one domain and sensible domain assignments.
|