Bayesian methods 1
Presenter: Gino Kpogbezan
When: Friday, July 15, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
An empirical Bayes approach to network recovery using external knowledge
Reconstruction of a high-dimensional network may benefit substantially from the inclusion of prior knowledge on the network topology. In the case of gene interaction networks such knowledge may come for instance from pathway repositories like KEGG, or be inferred from data of a pilot study. The Bayesian framework provides a natural means of including such prior knowledge. Previously we employed a Bayesian Structural Equation Model (BSEM), in combination with a variational algorithm and empirical Bayes estimation of hyperparameters. This method is computationally fast, and can outperform known competitors. Here we extend this approach to incorporate external data. In a simulation study we show that accurate prior data can greatly improve the reconstruction of the network, but need not harm the reconstruction if wrong. We demonstrate the benefits of the method in an analysis of gene expression data from GEO.
Bayesian methods 1
Presenter: Johannes Bogaards
When: Friday, July 15, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
A bivariate mixture model for natural antibody levels to human papillomavirus types 16 and 18: baseline estimates for monitoring the effects of immunization
Age-stratified serological surveys are widely used to study infectious disease dynamics, and provide a feasible and economic means to assess the population impact of vaccination against multi-strain pathogens. Post-vaccine monitoring programs for human papillomavirus (HPV) have been introduced in many countries, but HPV serology is still an underutilized tool, partly owing to the weak antibody response to HPV infection and the considerable potential for misclassification when considering multiple HPV types at once. Changes in antibody levels among non-vaccinated individuals could be employed to monitor herd effects of vaccination against HPV16 and -18, but inference requires an appropriate statistical model. The authors developed a four-component bivariate mixture model for jointly estimating HPV16 and -18 seroprevalence as a function of age from correlated antibody responses. The model was fitted to vaccine-type antibody concentrations as measured by a multiplex immunoassay in a large serological survey (3,875 females between 0 and 79 years of age) carried out in the Netherlands in 2006/07, before the introduction of preadolescent girls' vaccination. Parameters were estimated by posterior simulation with JAGS. We evaluated a suite of mixture models to describe the joint distribution of HPV16 and -18 antibody concentrations, differing with respect to correlation structures and number of marginal mixture components. The DIC was used for model selection; performance of the preferred model was assessed through simulation. Our analysis uncovered elevated antibody concentrations in doubly as compared to singly seropositive individuals, and a strong clustering of HPV16 and -18 seropositivity, particularly around the age of sexual debut. In addition, HPV16 and -18 antibody concentrations were strongly correlated in the seronegative and doubly seropositive mixture components. The joint analysis resulted in a more reliable classification of singly and doubly seropositive individuals than achieved by a combination of two univariate models, and suggested a higher pre-vaccine HPV16 seroprevalence than previously estimated. The bivariate mixture model provides valuable baseline estimates of vaccine-type seroprevalence and may prove useful in seroepidemiologic assessment of the herd effects from HPV vaccination.
Bayesian methods 1
Presenter: James Faulkner
When: Friday, July 15, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
Adaptive nonparametric smoothing for capture-recapture models
We extend a locally-adaptive nonparametric smoothing method known as the Bayesian trend filter for use in capture-recapture modeling. We use this method to estimate nonlinear temporal trends in survival and capture probabilities within a Cormack-Jolly-Seber capture-recapture modeling framework. This fully Bayesian approach places a horseshoe prior on the kth-order differences in the discretized latent trend function. This formulation allows adaptation to local changes in smoothness of a function, including abrupt changes or jumps, without compromising smoothness across the rest of the function. We use Hamiltonian Monte Carlo to approximate the posterior distribution of model parameters because this method provides superior performance in the presence of the high dimensionality and strong parameter correlations exhibited by our models. We use simulated data to assess the performance of the method, and we demonstrate its application with two real data sets.
Bayesian methods 1
Presenter: Sabine Hoffmann
When: Friday, July 15, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
Accounting for exposure uncertainty and effect modification using a Bayesian structural approach: Radon exposure and lung cancer mortality in a prospective cohort of uranium miners
Uncertainty in exposure assessment in occupational cohort studies can give rise to complex patterns of measurement error. It is well established that this exposure measurement error can have deleterious consequences for inference, including a loss in statistical power and biased point and interval estimates. In the French cohort of uranium miners, one is confronted with heteroscedastic measurement error, where both type and magnitude of error depend on period of exposure. We use a Bayesian structural approach, which simultaneously models classical and Berkson measurement error and the association between lung cancer mortality and radon exposure. It allows to account for both exposure and parameter uncertainty in a unique and coherent framework. It is of particular importance to use a well-fitting disease model, since all deviations from this model are considered as measurement error. We compare several disease models based on the inclusion of different effect modifying variables including period of exposure and time since exposure. Bayesian inference is conducted via an adaptive Metropolis-within-Gibbs algorithm implemented in Python. A simulation study suggests a substantial reduction of bias with such an approach. When applying it to the cohort, accounting for exposure uncertainty, one observes a slight increase in the risk estimate for lung cancer mortality associated with cumulated radon exposure. The most important effect modifying variable in the cohort is period of exposure with a risk estimate that is five times bigger for exposures received after 1955 than until 1955. This effect modification seems to result in an attenuation of the exposure risk relationship for high exposures, a phenomenon that has been reported previously. Future research is needed to account for further sources of uncertainty associated with the calculation of absorbed lung dose. In particular, this calculation of absorbed lung dose will allow us to disentangle the effects caused by radon and by other sources of ionizing radiation including long-lived radionuclides and gamma rays while accounting for dose uncertainty for all three exposures.
Bayesian methods 1
Presenter: Eva-Maria Huessler
When: Friday, July 15, 2016 Time: 11:00 AM - 12:30 PM
Room: Salon B Carson Hall (Level 2)
Session Synopsis:
A Bayesian Hierarchical Model for the Analysis of PAR-CLIP Data
microRNAs are small non-coding RNAs that play an important role in gene regulation, as they bind to target mRNAs to initiate translational repression and mRNA destabilization. Such targeted mRNA regions can be detected using PAR-CLIP (Photoactivatable-Ribonucleoside-Enhanced Crosslinking and Immunoprecipitation). The transition of T-to-C in the sequenced cDNA induced by this biotechnology helps to distinguish potential binding sites from noise. Nevertheless, detected T-to-C nucleotide exchanges can also be due to other reasons such as sequencing errors or single nucleotide polymorphisms (SNPs). Only few statistical methods have been developed to detect potential binding sites, but most of them do not account for errors and SNPs. One exception is wavClusteR [1], a mixture model based method, that accounts for SNPs and errors by employing information from other mutations than T-to-C. In this approach, a posterior distribution is determined for every single mutation position using a simple Bayesian model. We have extended this procedure to a fully Bayesian hierarchical model enabling the direct computation of a posterior distribution for all mutation positions via a MCMC algorithm. In our talk, we will present this Bayesian procedure, which is - to the best of our knowledge - the first method for the analysis of PAR-CLIP data that allows to incorporate additional information relevant for the biology of microRNA binding sites such as the mRNA region. Moreover, we will present an application of our method to a publicly available dataset and a simulation study in which the performance is compared to other procedures. References: [1] Sievers, C., Schlumpf, T., Sawarkar, R., Comoglio, F., Paro, R. (2012). Mixture models and wavelet transforms reveal high confidence RNA-protein interaction sites in MOV10 PAR-CLIP data. Nucleic Acids Research, gks697.
