English
Slovenščina
Seminar prof. Rubina
Ob prejemu častnega doktorata Univerze v Ljubljani Center za družboslovno informatiko in Katedra za družboslovno informatiko na FDV organizirata dva seminarja prof. Rubina, in sicer na temi manjkajoči podatki (3.12.) ter vzročnost (4.12.).
Oba seminarja bosta na FDV, predvidoma v senatni sobi.
Profesor Rubin je sicer eden največjih statistikov današnje dobe - podrobnosti >>. Njegova domača stran je tukaj>>.
Seminarja sta brezplačna, potrebna pa je prijava >>
Missing data (3.12.2012 ob 14.00 - 17.00)
The first part of the lecture will cover the basic concept of missing data, what constitutes missing data, patterns of missingness, and naive ways to deal with the problem of missing data. We will then consider reasons for missingness and the classification of these reasons into MCAR, MAR, and MNAR mechanisms. We will continue with a discussion of principled methods to address missingness, primarily maximum likelihood, usually implemented by EM, and Bayesian, usually implemented by MCMC as well as multiple imputation methods. Further, we will address why it is important to investigate the sensitivity of statistical conclusions to assumptions about the missingness mechanism. The second part of the lecture will present methods for displaying the sensitivity of statistical conclusions to these reasons with the use of graphical methods that rely on modern computing, in particular, a new method called “enhanced tipping point displays”.
Causality in Experiments and Observational Studies (4.12.2012 ob 16.00-19.00)
The first part of the lecture will carefully define causal effects using potential outcomes, and thereby view causal inference as a missing data problem. The second part will describe how to learn about causal effects in the simple setting of randomized experiments, using both Fisherian and Neymanian randomization-based methods, whose starting point is design. We will continue with a discussion on how to design observational studies to approximate randomized experiments, primarily using propensity score methods. In this part we will also address Bayesian or direct likelihood approaches which are needed to deal effectively with complex situations, even in randomized experiments with complications, such as noncompliance, unplanned missing data, or partially defined outcomes, such as hourly wages for the unemployed.