 |
Postdoctoral research at Imperial College |
Institute: Imperial College of Science, Technology and Medicine
Research department: Epidemiology and Public Health (EPH)
Research team: Biostatistics group
Place: St Mary's Campus, Norfolk Place, London (UK)
Collaborations : Sylvia Richardson (Imperial College, Department EPH)
Juan Jose Abellan (CIBER Epidemiologia y Salud Publica, Spain)
the Veterinary Laboratories Agency (UK)
Summary of project 1: Recent advances in disease mapping have focused on jointly analysing the spatial or spatio-temporal variations of risks of several diseases with common risk factors from a unique data source. The goal of this project was to build and implement several Bayesian shared spatial component models to pool together and borrow strength across several data sources informing about a common underlying disease risk surface. The proposed models were applied to the spatial analysis of risk of scrapie infection in Wales (UK).
Keywords: Bayesian hierarchical modelling, conditional autoregressive, disease mapping, risk assessment, MCMC algorithms, multiple health indicators, multivariate analysis
Summary of project 2: In collaboration with Lea Fortunato (Imperial College), I realized a comparative study of the new Bayesian inference software INLA (Integrated Nested Laplace Approximation) proposed by Havard Rue and the Bayesian user-friendly software WinBUGS from a large amount of geo-referenced cancers data in UK. Our ultimate goal was to provide maps describing the spatial variations of the risk of 30 different types of cancer in UK. I took advantage of the extremely short computation time provided by INLA to make a large sensitivity analysis to prior and modelling assumptions for each type of cancer.
Keywords: disease mapping, Gaussian hidden markov random fields, Laplace Approximation, prior sensitivity, risk assessment
|
|
Postdoctoral research at INSERM |
Institute: INSERM
Research Unit: Joint research unit U822 INSERM-INED-PARIS XI
Research team: Epidemiology of the reproduction
Place : Hôpital de Bicêtre (Le Kremlin-Bicêtre, next to Paris, France)
Collaborations with :
Jean Bouyer (INED-INSERM-PARISXI UMR 822)
Elise de la Rochebrochard (INED-INSERM-PARIS XI UMR 822)
Michel Chavance (INSERM Unit 780 "Epidemiology & Biostatistics")
Summary of the project: In most of medical studies, modelling longitudinal data often is complicated by missing data due to the attrition of cohorts. Generally speaking, these study outputs don't occur completely at random. More presisely, they depend on observed and unobserved response variables : they are said to be informative. The goal of this work is to develop and implement Bayesian shared random-effects models to analyze binary longitudinal data with informative censoring. This class of models defines an indirect stochastical link between the response process and the missing-data mechanism through one or several shared random effects. Up to our knowledge, this class of statistical models is less used in practice and in the Bayesian context. This work is motivated by an original application in reproductive health. The main goal of the study is to estimate the average probability for a woman to give birth after an In-Vitro Fertilization (IVF) treatment in an hypothetical population in which no woman could drop out during her medical treatment. Moreover, another goal is to be able to quantify the degree of the association between the binary response process (i.e., giving birth or not after an IVF attempt) and the drop-out behavior of women during their IVF program.
Keywords: attrition, binary longitudinal data, generalized linear mixted-effects models, informative censoring, missing data, multivariate modeling
|
|
Research associate at INRA |
Institute: INRA
Research department: Applied Mathematics and Informatics
Research team: Met@risk (Methodologies of food risk analysis)
Place : AgroParisTech (Paris, France)
Collaboration: Isabelle Albert (INRA, research team Met@risk)
Context: the research program Quant'HACCP funded by the French Research National agency in collaboration with: AFSSA,ENVA, CEMAGREF, AGROPARISTECH, IFIP......
(More details on the web site: http://www.quanthaccp.fr/index.html)
Summary of my research works : To describe the transmission dynamics of a population of pathogens from "farm to table", the classical statistical approach consists in building a stochastic exposure model that covers the food processing and storage chain then using Monte Carlo simulations to perform a quantitative microbiological risk analysis on the model's outputs given the inputs. Such an approach takes into account the different sources of variability and uncertainty inherent to the model’s inputs from prior elicitation based on expert opinion, external data or literature. In this work, we will first intend to improve this classical approach by making use of additional available data corresponding to the model’s outputs (e.g. the contamination at the end of the domestic storage). The main idea is to build the overall model as a Bayesian network by combining a stochastic core model based on expert knowledge with a probabilistic data model (or sampling model) taking into account the information brought by the data collected at the end of the food processing chain.
Setting critical limits at control points defined throughout the food processing chain is supposed to be based on a quantitative justification. However, in practice, few quantitative tools are used to achieve this. In a second step, we will intend to develop an efficient statistical approach to define these critical limits from food safety objectives. In other words, we will aim at developing then validating a thresholding method on the parameters’ joint distribution (i.e., the inputs) according to constraints defined on the model outputs.
Keywords: Bayes networks, quantitative risk analysis, thresholding method
|
