Il 12 ottobre 2018 avrà luogo la presentazione del primo gruppo di tesi di dottorato di ricerca in Scienze Statistiche, curriculum Statistica metodologica.
Ore 9.30-10.10
W.J. Radermacher
Title: Science, Statistics, Society. Towards Official Statistics 4.0
Abstract: Riportato alla fine della mail
Ore 10.10-10.20
Discussione
Ore 10.20-11.00
M. Stefanucci
Titolo: Statistical Methods for Classification of Partially Observed Functional Data
Abstract: Riportato alla fine della mail
Ore 11.00-11.10
Discussione
Ore 11.10-11.30: Pausa
Ore 11.30-12.10
S. D'Angelo
Titolo: Latent space models for multidimensional network data
Abstract: Riportato alla fine della mail
Ore 12.10-12.20
Discussione
Ore 12.20-13.00
T. Padellini
Titolo: Interpretable statistics for complex modelling: quantile and topological learning
Abstract: Riportato alla fine della mail
Ore 13.00-13.10
Discussione
ABSTRACT DELLE PRESENTAZIONI
Ore 9.30-10.10
W.J. Radermacher
Title: Science, Statistics, Society. Towards Official Statistics 4.0
Abstract
The term ‘statistics’ is used differently; it can refer to a science, a certain kind of information or institutions. This work is concerned neither with statistics in general nor with the history of theoretical statistics. Rather, the goal is to describe the status quo for a particular area of application, namely ‘official statistics’.
Central to this work is the quality of statistical information. Statistics can only develop a positive enlightenment effect on the condition that their quality is trusted. To ensure long-term trust in statistics, it is necessary to deal with questions of knowledge, quantification and the function of facts in the social debate. How can we know that we know what we know (or do not know)? The more concrete an answer that can be given to such questions, the more possible it will be to protect statistics against inappropriate expectations and to address false criticism.
When one uses the term ‘official statistics’, one deals with the problem that again different meanings are possible, namely the institution (the statistical office), the results (statistical information) and, of course, the processes (the surveys). But of course, one associates with the notion of official statistics first that it deals with social and economic issues, and more recently also with ecological ones. How many people live in a country, how much is produced, what about work, health and education? We encounter these and related topics daily in the media, in political discussions and decisions. From them we expect a solid quality; we must trust them.
Official statistics play a fundamental role in modern societies: they are an essential basis for policies, they support business decisions and they allow citizens to evaluate the progress made. But the power of statistical knowledge also poses dangers. From a cognitive tool that can emancipate and promote participation, it can transform itself into a true technocratic tyrant, to varying degrees, behind evidence-based decision-making and mainstream management ideologies.
At the end of a period of automation, official statistics today are considerably more efficient, faster, and better in quality compared to the 1980s. But what has now begun with a comprehensive digitisation of our world demands in many ways a more radical and innovative strategy. In addition, globalisation calls for new answers that cannot always be achieved with continuous development, but that will be accompanied by disruptions and more radical changes.
In preparation for deriving conclusions and strategies for official statistics development in the near future, it is helpful, indeed necessary, to take a closer look at the driving forces: the scientific background of official statistics as well as episodes from the history of the past 200 years, insofar as they are relevant to the understanding of ‘Statistics 4.0’
Ore 10.20-11.00
M. Stefanucci
Titolo: Statistical Methods for Classification of Partially Observed Functional Data
Abstract
This thesis is devoted to the development of new methodologies for the classification of partially observed functional data. Functional Data Analysis is nowadays one of the most active area of research in statistics. It deals mostly with data coming from technical machineries and digital instruments, treating data as functions. Classification of this kind of data is still an open problem and there are several available approches in the literature. Unfortunately, essentially none of them is directly applicable when the data are partially observed, i.e. exhibit some missing parts. The aim of this work is to provide new insights and proposals for discrimination of fragmented functional data. The theory we develop is strongly supported by extensive simulations and all the methods are illustrated on a real medical dataset, on which we outperform previous classification attempts.
Ore 11.30-12.10
S. D'Angelo
Titolo: Latent space models for multidimensional network data
Abstract
Network data is any relational data recorded among a group of individuals, the nodes. When multiple relations are recorded among the same set of nodes, a more complex object arises, called “multidimensional network”, or “multiplex”. In the past, statistical analysis of networks mainly focused on single-relation network data. Only in recent years, statistical models specifically tailored for multiplex data begun to be developed. In this context, a few works extended the latent space modeling framework to multiplex data, where it is postulated that nodes have latent positions in a p−dimensional Euclidean space. When considering multidimensional networks, these unknown positions are assumed to be relevant in determining the observed multiplex structure. Also, they may help capturing shared node similarities.
This thesis develops novel latent space models for multidimensional networks, and tackle different features that observed multiplex data may present. In a first work, we focus on modeling multiplex data that show an high degree of symmetry between networks. Here, different networks may be assumed to be realizations of a single underlying phenomenon and such replicated measures may be used to model overall similarities between the nodes. A second work introduces a class of latent space models with node-specific effects, able to deal with different degrees of heterogeneity within and between networks in multiplex data. A third work addresses clustering of the nodes in the latent space, where clusters correspond to communities of nodes in the multiplex.
A Bayesian hierarchical modeling approach is adopted and all models are tested via different simulation studies. Further, their performances are illustrated via real world data examples.
Ore 12.20-13.00
T. Padellini
Titolo: Interpretable statistics for complex modelling: quantile and topological learning
Abstract
As the complexity of our data increased exponentially in the last decades, so has our need for concise but highly expressive and interpretable feature sets. This thesis revolves around two paradigms to approach this quest for insights: topological and quantile learning.
In the first part we focus on parametric models, where the problem of interpretability can be seen as a "parametrisation selection". We introduce a quantile-centric parametrisation and we show the advantages of our proposal in the context of regression, where it allows to bridge the gap between classical generalised linear (mixed) models and increasingly popular quantile methods.
The second part of the thesis, concerned with topological learning, tackles the problem from a non-parametric perspective. As topology can be thought of as a way of characterising data in terms of their connectivity structure, it allows to represent complex and possibly high dimensional through few features, such as the number of connected components, loops and voids. We illustrate how the emerging branch of statistics devoted to recovering topological structures in the data, Topological Data Analysis, can be exploited both for exploratory and inferential purposes with a special emphasis on kernels that preserve the topological information in the data.
Finally, we show with an application on brain imaging data how these two approaches can borrow strength from one another to profile patients with autism spectrum conditions.