Paola Vicard: Bayesian networks in official statistics

Paola Vicard (Universita di Roma Tre) will talk about

Bayesian networks in official statistics

Abstract

Statistical analyses can be particularly complex when are referred to surveys and databases produced by a National Institute of Statistics. The complexity is mainly due to: high number of surveys carried out by the institute, sampling design complexity, high number of variables and huge sample size. In this context it can be useful to analyse and exploit the dependence structures. Bayesian networks are multivariate statistical models able to represent and manage complex dependence structures. The theoretical setting of BNs and of graphical models is the basis for developing methods for efficiently representing and managing survey systems. A known (or previously estimated) dependence structure can help: in estimators computation (with the sampling design either explicitly or implicitly modelled); when coherence constraints among different surveys must be fulfilled; in integrating different sample surveys (in terms of their joint distribution); in updating the estimate of a joint distribution once new knowledge on a variable marginal distribution occurs (survey weights poststratification is a special case); in missing data imputation. In this framework, also the structural learning problem can be considered. Here we focus on the score+search structural learning algorithm, for which the most likely structure, given the observed data, will be identified optimizing an objective function (typically a penalized likelihood), by means of optimization methods that, in this case, will be given by the greedy search and the genetic algorithms ones.

Published Apr. 13, 2012 3:49 PM - Last modified June 5, 2012 10:11 AM