18 Maggio 2020 ore 12.30 Webinar

Prof. Panagiotis TSIAMYRTZIS
Depat. of Mechanical Engineering, Politecnico di Milano

Predictive Control Charts (PCC). A Bayesian Approach in Online Monitoring of Short Runs

Abstract
In Statistical Process Control and Monitoring (SPC/M), performing online monitoring for short horizon of phase I data is a challenging, though cost effective benefit. Self-starting methods attempt to address this issue adopting a hybrid scheme that executes calibration and monitoring simultaneously. In this work, we propose a Bayesian alternative that will utilize prior information and possible historical data (via power priors), offering a head-start in online monitoring, putting emphasis on outlier detection.

Charting will be based on the predictive distribution and the methodological framework will be derived in a general way, to facilitate discrete and continuous data from any distribution that belongs to the regular exponential family (with Normal, Poisson and Binomial being the most representative). Being in the Bayesian arena, we will be able to not only perform process monitoring, but also draw online inference regarding the unknown process parameter(s).

An extended simulation study will evaluate the proposed methodology against frequentist-based competitors and will cover topics regarding prior sensitivity and model misspecification robustness. A continuous and a discrete real data set will illustrate its use in practice. Short production runs and online phase I monitoring are among the best candidates to benefit from the developed methodology.

23 Marzo 2020 ore 12,00     EVENTO ANNULLATO
 

Prof.ssa Francesca CHIAROMONTE
The Pennsylvania State University
 

"Oms" perspectives on Childhood Obesity
 

Abstract:
Childhood obesity is a growing epidemic with an enormous societal impact. As part of an ongoing collaboration with the INSIGHT and SIBSIGHT projects led by Dr. Ian Paul and colleagues at Penn State Hershey Medical Center, we focused on analyzing a variety of “omics” data collected from children and their mothers. These included gut and oral microbiota, genetic variants, and fecal metabolomes – along with important clinical, behavioral, and environmental covariates. Notwithstanding the relatively small sample size of the study, we were able to leverage a number of statistical techniques for handling high-dimensional predictors and functional response variables (in our case, the growth curves of children sampled over the first three years after birth) to identify significant “omics” risk signatures associated with weight gain in early life. In this talk I will briefly survey our analysis approaches and results.

This is joint work with Drs. Sarah Craig and Kateryna Makova (Biology), Anna Kenney and Matthew Reimherr (Statistics) and many other colleagues across the Pennsylvania State University.

21 Febbraio 2020 ore 14,15

Prof.ssa Laura VENTURA
Università degli Studi di Padova

Bayesian posterior distributions from estimating functions

Abstract:

We discuss the use of estimating equations in order to compute a posterior distribution in two frameworks.

1) Standard Bayesian analyses can be difficult to perform when the full likelihood, and consequently the full posterior distribution, is too complex and difficult to specify or if robustness with respect to data or to model misspecifications is required. In these situations, we suggest to resort to a posterior distribution for the parameter of interest based on unbiased estimating equations, which include as special instances M-estimating functions and proper scoring rules. We show that the posterior distribution can be derived analytically or by approximate Bayesian computation methods.

2) Consider Bayesian inference when the prior distribution is known only through its first derivative (typically with multidimensional parameters and objective priors, e.g. matching priors). In this case it is only possibile to evaluate the first derivative of the log-posterior. In this framework, we propose two methods to derive a posterior distribution for a parameter of interest using a Rao score statistic or a local Taylor approximation of the log-posterior and its first derivative.

12 febbraio 2020 ore 11.30

Ernst C. WIT
Università della Svizzera italiana, Lugano, Switzerland
 

High-dimensional inference in graphical models

Abstract:

The modern paradigm for biological, economic, sociological processes is an interconnected, complex system where various parts interact in complex ways with each other. This process can often be modelled either by means of a random graph or a graphical model. In this talk we will focus on the latter.

Our aim is to describe general inference methods that can deal with high-dimensional systems. We show a Bayesian approach incorporated in the R-package BDgraph (with R. Mohammadi) and a frequentist approach (with A. Abbruzzo and I. Vujacic), with particular biological applications, that we implemented in netgwas (with P. Behrouzi) and rMAGMA (with A. Cougoul).