Search

We propose a tool for exploring the number of clusters based on pivotal methods and consensus clustering. K-means algorithm is used to learn the pairwise similarity via the co-occurrence of points in multiple partitions of the data. This similarity can be used to investigate the number of groups and detect arbitrary shaped clusters. Different criteria for identifying the pivots are discussed, as well as preliminary results concerning the selection of the optimal number of clusters.

When modeling the relationship between a response and some continuous covariates, assuming linearity may be too restrictive in many contexts. Naive solutions to overcome this limitation, such as categorisation of the predictor, have well-known drawbacks. A viable alternative is represented by spline functions. In epidemiological studies, the number and position of knots usually have an important meaning. Therefore, special attention should be posed to techniques that allow to choose the number and position of knots. The aim of the present work is to: 1. introduce a two-step Bayesian procedure within the semiparametric generalised linear model framework to be applied in epidemiological studies where the effect of a continuous exposure on risk is under investigation; 2. show how this framework is applied in a bivariate context where the aim is to modeling the joint effect of intensity and duration of alcohol drinking in cancer of the oral cavity.

Spline functions, defined as piecewise polynomials with a fixed degree, whose joint points are called knots, are highly flexible tools to modeling non-linearity between a response and some continuous covariates. In epidemiological studies, the number and position of knots usually have an important meaning. Therefore, special attention should be posed to techniques that allow to choose the number and position of knots. Here, we will follow one of the most recent approaches to variable selection in a Bayesian context. Estimating the positions of the knots is not easy and, for a fixed degree, regression coefficients and locations of knots have to be estimated simultaneously, turning the estimation into a non-linear optimisation problem. The aim of the present work is to: 1. introduce a two-step Bayesian procedure within the semiparametric generalised linear model framework, to be applied in epidemiological studies where the effect of a continuous exposure on risk is under investigation; 2. show how this framework is applied in a bivariate context, where the aim is to modeling the joint effect of intensity and duration of alcohol drinking in cancer of the oral cavity.

In modelling football outcomes, scores’ data are regularly used for the estimation of the attack and the defence strength of each team. However, these teams’ abilities are quite complex and are correlated with many quantities inherent to the game. Additional available information, relevant for their estimation, are shots, both made and conceded. For such a reason, we propose a hierarchical model that incorporates this information in three stages for each game and each team: number of scores, number of shots on target and number of total shots. We fit the model on English Premier League data and obtained predictions for future matches.

A Stan model for soccer shooting ability., Code and data available at github.com/stan-dev/stancon_talks

Obiettivo del volume è introdurre il modello di regressione lineare, con cenni alle principali estensioni.

We review the recent debate on the lack of reliability of scientific results and its connections to the statistical methodologies at the core of the discovery paradigm. Contributions to the debate include recommendations for paradigm changes as well as methods to quantify the reliability of the present paradigm. We review the main proposals within the two approaches and we propose a new modeling strategy for assessment purposes to exploit data from replication experiments.

Recently, the Italian Statistical Agency Istat faced the problem of defining the new SLL (Local labour Systems) considering the data on commuting flows between municipalities collected in the last Italian Census (2011) by using a nonhierarchical agglomerative clustering procedure. We propose instead a model based clustering technique that directly accounts for network relations between municipalities and their position in geographical space. The model is based on latent and manifest geographical coordinates, a gaussian mixture distribution for the coordinates is considered and allows us to obtain a clustering of the municipalities based on the flows (having discounted for the population sizes). The model is estimated on commuting flows between municipalities in Friuli-Venezia Giulia region using the Bayesian approach. Results are compared with those recently presented by Istat.

We consider pre-electoral polls for Italian general elections of 2013 with the aim of shedding light on pollsters behavior. We compare vote share prediction variability across parties and pollsters and model how this variability changes over time. Estimates of a Bayesian hierarchical model showed that a large portion of the total variability of vote share predictions is explained by the so-called house effect. Furthermore we noted that the variability of vote share predictions slightly reduces over time as the election day approaches.

Nel presente lavoro si discutono alcune criticità insite nel'impiego delle valutazioni sulla qualità della didattica percepita dagli studenti: la confrontabilità tra le valutazioni di insegnamenti diversi e l'interpretabilità di tali confronti come confronti di abilita tra docenti. Per quanto riguarda il primo aspetto si mostrera, anche utilizzando un insieme di dati sulla qualità della didattica percepita dagli studenti presso un grande ateneo, come l'uso di un approccio gerarchico Bayesiano permetta di superare alcune situazioni paradossali e di pervenire a una pi\u corretta e coerente sintesi della valutazione della qualita della didattica per singoli docenti, per gruppi di insegnamenti oppure corsi di laurea. Per quanto attiene all'interpretabilit\`a in termini di abilit\`a dei docenti si discute di quali dovrebbero essere le caratteristiche della rilevazione per consentire un'interpretazione causale in cui migliore valutazione \`e causata da migliore docente.

Despite being a standard tool for data analysis in many scientific fields, statistical testing has always been associated to criticism and controversies. Recently, the debate on the role of statistical testing in science has been particularly vivid and coupled with the debate on the lack of reproducibility of scientific results in many disciplines. We review some recent critics and purported consequences of statistical testing use and abuse and discuss proposed remedies, which entails either improving testing or abandoning it for alternative, more appropriate, strategies.

Over the period 2001-2005 Italy experienced an expansion of enrollment in higher education benefiting from the Bologna Process. This paper evaluates the effects of this policy on academic performance: we compare the performances of students who switch from pre-reform university courses to post-reform ones with that of students who stay in pre-reform university courses. The aim of the paper is to estimate the difference in graduation probability, allowing for the changing university regime (the treatment), job opportunities and other characteristics of the students.We considered a sample of 25,866 undergraduate students enrolled at University of Trieste, which is reduced to a balanced sample of 1,020 units by matching techniques. Estimates of a logistic models on the balanced sample suggest that the treatment effect differs between the “old” and the young students: the treatment appearing beneficial for the younger ones only.

It is widely known that pre-electoral polls often suffer from non-sampling errors which pollsters try to compensate in final estimates by means of diverse ad hoc adjustments, thus leading to the well-known house effects. We analyze vote share predictions from election polls in Italy in 2006, 2008 and 2013 in order to investigate the relative role of house effects on their variability. We are able to confirm that the variability due to house effect is the most important source of variation

In this paper we study the factors influencing dropout in the second year of the bachelor programs at the University of Trieste. We control for the different withdrawal causes. In doing so, we apply the Lasso method for the selection of the covariates determining the dropout event described in Tibshirani (1996). We use administrative data on 23,333 undergraduates students enrolled in three-years bachelor programs from 2003 to 2011 and perform a binomial GLM model with logistic link. Our aim is to improve the general understanding of the students’ withdrawing focusing on personal characteristics of students and on institutional aspects of the university. We demonstrate that a high secondary school final mark and a low individual students’ performance increase significantly the risk of dropping out, the same is true for the choice of a "weak" degree. In general, the low entry barriers permit to the students coming from lower socioeconomic background to limit dropout along the whole academic path. Our results reveal, also, that obtaining a job in the first year of study or being "old" or residing in the region, will increase the dropout probability. Thus, to promote degree completion and decreasing dropout and "parking time", universities should enlarge the tools provided to students in order to manage failure.

We consider a model based clustering technique that directly accounts for network relations between subjects and their position in geographical space. A natural framework to deal with this issue is social network analysis, where municipalities are the nodes and the relationship between them is measured by the commuting flows. In our model the flows are explained by the distances between the nodes in a three-dimensional space where two coordinates are the actual geographical coordinates and the third one is a latent variable. The model is completed by specifying a gaussian mixture distribution for the coordinates: this model component allows us to obtain a clustering of the municipalities based on the flows (having discounted for the populations).