The goal of this project is to develop improved statistical methods for the analysis of data from human studies, with an emphasis on epidemiologic and genetic studies. A focus has been on improving the flexibility and sensitivity of methods for assessing complex environmental and genetic effects on health outcomes, including outcomes that are multivariate and subject to complex censoring mechanisms. Areas of focus include (1) developing methods for high-dimensional dependent predictors, such as correlated exposure variables in environmental epidemiology and single nucleotide polymorphisms (SNPs) in genetic studies; (2) developing flexible statistical models that allow heterogeneity in the effect of environmental exposures, while allow interactions with genetic factors and other predictors; (3) developing methods for assessing effects of time-varying predictors; (4) improving methods for accommodating model uncertainty.? ? In the first area, we have developed high-dimensional variable selection methods that rely on Dirichlet process mixture priors to adaptively allocate predictors to clusters defined by the magnitude of the health effect. We have applied these approaches for selection and clustering of polymorphisms in functionally-related genes and for identifying important environmental exposures from among a set of highly-correlated candidates within a mixture. In the second area, we have developed a general method for Bayesian density regression and classification based on adaptive mixtures of linear and logistic regression models. These models are highly flexible due to the ability to change the regression coefficients for different values of the predictors, an idea related to splines, but more flexible in allowing the whole response distribution to change instead of just the mean. The approach has been applied to several applications with good results, including data from comet assay studies assessing genetic and environmental predictors of DNA repair rates. In the third area, we have developed a new semiparametric modeling framework, referred to as a joint functional Dirichlet process (JFDP). The JFDP automatically clusters functional predictors, while using the cluster status to nonparametrically predict the joint distribution of one or more health outcomes. For example, we have used this approach to study water quality effects on the joint distribution of gestational age at delivery and birth weight. In the fourth area, we have developed methods for accommodating uncertainty in random effects models for longitudinal data, letting both the predictors to be included and the distribution of their random effects be unknown.
| Dunson, David B; Park, Ju-Hyun (2008) Kernel stick-breaking processes. Biometrika 95:307-323 |
| Pennell, Michael L; Dunson, David B (2007) Fitting semiparametric random effects models to large data sets. Biostatistics 8:821-34 |
| Kinney, Satkartar K; Dunson, David B (2007) Fixed and random effects selection in linear and logistic models. Biometrics 63:690-8 |
| Dunson, David B (2007) Bayesian methods for latent trait modelling of longitudinal data. Stat Methods Med Res 16:399-415 |
| MacLehose, Richard F; Dunson, David B; Herring, Amy H et al. (2007) Bayesian methods for highly correlated exposure data. Epidemiology 18:199-207 |
| Dunson, David B (2006) Bayesian dynamic modeling of latent trait distributions. Biostatistics 7:551-68 |
| Baird, Donna D; Kesner, James S; Dunson, David B (2006) Luteinizing hormone in premenopausal women may stimulate uterine leiomyomata development. J Soc Gynecol Investig 13:130-5 |
| Longnecker, Matthew P; Klebanoff, Mark A; Dunson, David B et al. (2005) Maternal serum level of the DDT metabolite DDE in relation to fetal loss in previous pregnancies. Environ Res 97:127-33 |
| Dunson, David B; Herring, Amy H (2005) Bayesian model selection and averaging in additive and proportional hazards models. Lifetime Data Anal 11:213-32 |
| Dunson, David B; Herring, Amy H (2005) Bayesian latent variable models for mixed discrete outcomes. Biostatistics 6:11-25 |
Showing the most recent 10 out of 27 publications
