Brown University Center for Statistical Sciences Seminar
Simon Fraser University | |
Abstract: With concerns of bioterrorism, the advent of new epidemics that spread with person-to-person contact, such as SARS, and the rapid growth of on-line social networking websites, there is currently great interest in building statistical models that emulate social networks. Stochastic network models can provide insight into social interactions and increase understanding of dynamic processes that evolve through society. A major challenge in developing any stochastic social network model is the fact that social connections tend to exhibit unique inherent dependencies. For example, they tend to show a lot of clustering and transitive behavior, heuristically described as "a friend of a friend is a friend." It might be reasonable to expect that covariate similarities, or "closeness" in social space, should somehow be related to the probability of connection for some social network data. The relationship between covariates and relations is likely to be complex, however, and may in fact be different in different regions of the covariate space. Here, we present a new socio-spatial process model that smoothes the relationship between covariates and connections in a sample network using relatively few parameters, so the probabilities of connection for a population can be inferred and likely social network structures generated. Having a predictive social network model is an important step toward the exploration of disease transmission models that depend on an underlying social network.
Candidate for Assistant Professor in the Biostatistics Section of the Program in Public Health.
Brown University Center for Statistical Sciences Seminar
Institute for Statistics and Decision Sciences, Duke University | |
(Refreshments beginning at 3:15 p.m.) |
Abstract: In multi-center studies, subjects in different centers may have different outcome distributions. This discussion is motivated by the problem of nonparametric modeling of these distributions, borrowing information across centers while also allowing centers to be clustered. Starting with a stick-breaking representation of the Dirichlet process (DP), we replace the random atoms with random probability measures drawn from a regular DP. This results in a nested Dirichlet process (nDP) prior, which can be placed on the collection of distributions for the different centers, with centers drawn from the same higher level component automatically clustered together. Theoretical properties are discussed, and an efficient MCMC algorithm is developed for computation. The methods are illustrated using a simulation study, an application to quality of care in US hospitals and an application to functional clustering in oceanography.
Candidate for Assistant Professor in the Biostatistics Section of the Program in Public Health.
<--- 2007 Index