Seminars


2026


Robust clustering based on trimming
January 26, 2026, 12:00
Clustering is one of the most widely used unsupervised learning techniques. However, it is well known that outliers can severely affect the performance of standard clustering methods. Clustered outliers, in particular, can be especially detrimental even for robust statistical procedures. This motivates the combination of concepts from Robust Statistics and Cluster Analysis to handle both clusters and outliers simultaneously through robust clustering approaches. Outliers may arise from errors in data collection, but they can also reflect genuine anomalies of significant scientific or commercial interest. Therefore, anomaly detection becomes a goal in itself, and the use of robust clustering techniques for their automated identification in heterogeneous datasets can be a very useful tool. Among existing robust clustering methods, we focus on those based on trimming. Trimming provides an intuitive framework: a small fraction of the most extreme observations is removed, and standard clustering methods are then applied to the remaining data. When combined with suitable constraints on cluster dispersion parameters, trimming-based approaches have demonstrated strong performance and can be efficiently implemented using available algorithms. This seminar will start from the fundamental principles of robustness and provide a guided overview of various robust model-based clustering techniques across different scenarios and frameworks. It will cover the necessary algorithms for implementation and discuss solutions to key challenges, including those arising from increasing dimensionality.

2025


A gentle statistical introduction to domain adaptation
December 5, 2025, 12:00
In the statistical learning framework, we assume that the learning set is an unbiased sample of the population of interest and the test set is used to assess the fitted model. Anyway, in real-world applications, the available data for the learning set can result to be a biased sample of the target population and therefore a classifier trained on these data may result to be quite inadequate. In other words, there is a “shift" between the two domains. There are many reasons that can explain this dissimilarity: for instance, collecting new labeled data is often time-consuming, costly, or even infeasible, especially when the statistical properties of the population evolve over time. In some other cases, we are provided with a large amount of unlabeled data, but only a small amount of labeled data. In the machine learning literature, the area dealing with this kind of problem is called Domain Adaptation. In this seminar, we present main ideas of the Domain adaptation in a statistical setting. Numerical studies will be further presented based on both simulated and real datasets.
Causal Inference when Intervention Units and Outcome Units Differ
November 28, 2025, 12:00
We study causal inference in settings characterized by interference with a bipartite structure. There are two distinct sets of units: intervention units to which an intervention can be applied and outcome units on which the outcome of interest can be measured. Outcome units may be affected by interventions on some, but not all, intervention units, as captured by a bipartite graph. Examples of this setting can be found in analyses of the impact of pollution abatement in plants on health outcomes for individuals, or the effect of transportation network expansions on regional economic activity. We introduce and discuss a variety of old and new causal estimands for these bipartite settings. We do not impose restrictions on the functional form of the exposure mapping and the potential outcomes, thus allowing for heterogeneity, non-linearity, non-additivity, and potential interactions in treatment effects. We propose unbiased weighting estimators for these estimands from a design-based perspective, based on the knowledge of the bipartite network under general experimental designs. We derive their variance and prove consistency for increasing number of outcome units. Using the Chinese high-speed rail construction study, analyzed in Borusyak and Hull [2023], we discuss non-trivial positivity violations that depend on the estimands, the adopted experimental design, and the structure of the bipartite graph.
Bayesian network propensity score estimation for testing causal effect in binary data
November 21, 2025, 12:00
Estimating treatment effects in real-world data is challenging due to potential bias arising from non-random treatment assignment and confounding variables. The propensity score (PS) is commonly used to correct for such bias, but its accuracy depends on properly modelling the treatment–covariate relationships. This study proposes estimating PS via Bayesian Networks (BNs), offering a flexible and often superior alternative to logistic regression, especially in complex dependency settings. The BN-based PS is applied to construct two estimators of the Average Treatment Effect (ATE): the Horvitz-Thompson (HT) and Hajek (H) types. When the PS model is correctly specified, both are asymptotically equivalent; under misspecification, the H-type performs better. Extensive simulations confirm the robustness and effectiveness of the proposed approach.
Bayesian hierarchical models for daily temperatures: means, quantiles, and record-breaking events
November 11, 2025, 12:00
This seminar presents a comprehensive Bayesian hierarchical framework for modeling daily temperatures, integrating mean, quantile, and record-breaking analyses across space and time. We first introduce a spatio-temporal mean model for daily maximum temperatures, incorporating two temporal scales, explicit autoregressive dependence, fixed effects, and multiple random effects to capture spatial variability. Building on this, we develop a mixed effects quantile regression model with asymmetric Laplace errors to explore climate change across quantiles, enabling marginal quantile inference from conditional autoregressive structures and revealing pronounced spatial-quantile heterogeneity in climate signals. Next, we introduce a new framework for analyzing high-temperature events, proposing hierarchical models for both univariate and bivariate record-breaking temperatures. The univariate model employs a logistic regression formulation with an explicit long-term trend and strong daily spatial random effects to analyze the occurrence of calendar-day records across years, allowing inference on the number, spatial distribution, and temporal evolution of record-breaking events under climate change. The bivariate model uses a probit regression to jointly model maximum and minimum temperature records, capturing spatial and temporal dependence through anisotropic coregionalized Gaussian processes and revealing correlated yet distinct patterns of record-breaking behavior. All models leverage Gaussian latent representations for closed form Gibbs sampling and provide spatial predictions at unobserved locations. Applications to long-term datasets from Spain illustrate trends in daily temperatures, quantile-specific climate change, and the increasing occurrence of record-breaking temperatures, offering new tools for the statistical analysis of environmental extremes.
From PoS to u-PoS: a journey through the realm of Probability of success
November 7, 2025, 12:00
The Probability of success (PoS) of a trial is conventionally defined as the expected value of the power function of a test with respect to a design prior assigned to the parameter under scrutiny. Even though this quantity is widely used in clinical trials for experimental design, the definition of probability of success is not univocal. In this presentation we review and compare the main types of PoS; specifically, we focus on a unifying, decision-theoretic approach that yields a new type of PoS as the expected utility of the trial (u-PoS), that is the expected probability of making the correct choice between two hypotheses. We base our comparisons on properties of the probability distributions of the power-related random variables associated to these definitions. Our conclusion is that u-PoS shows a conceptual advantage over previous versions of PoS; moreover, when suitable design priors are used, it may produce smaller optimal sample sizes with respect to its competitors.
Bayesian Regression Factor Model for Multivariate Causal Effect
October 31, 2025, 12:00
In the context of causal inference, the study of causal effects for multivariate potential outcomes remains relatively underexplored. This gap largely stems from the inherent difficulty of quantifying the overall causal impact of a treatment on correlated outcomes and understanding how these effects vary across different outcomes. Nevertheless, this research question is critically important in fields such as environmental epidemiology, where one may seek to assess the causal relationship between air pollution regulations and the concentrations of multiple pollutants, or between air pollution exposure and hospitalizations for various diseases. To address this challenge, we leverage a Bayesian factor regression model to identify latent, treatment-specific factors that capture causal effects among correlated multivariate outcomes. We propose a methodology that (i) introduces novel causal estimands within a general framework for multivariate outcomes, and (ii) develops a multi-treatment Bayesian factor regression model that enables the identification and characterization of causal latent effects. The innovative use of the dependent Dirichlet process as the distribution for the factor scores further allows us to handle missing data through a principled, awareness-driven, and fair imputation mechanism. The performance of the proposed method is demonstrated through both simulation studies and real-world applications in environmental epidemiology.
La statistica come arte critica
October 24, 2025, 12:00
La statistica occupa un ruolo centrale nel comprendere e gestire la complessità del mondo contemporaneo, ben oltre i confini delle scienze dure: si configura sempre più come una “disciplina critica”, capace non solo di analizzare dati, ma di mettere in discussione le categorie stesse con cui interpretiamo la realtà. Lungi dall’essere una semplice tecnica neutra, la statistica nasce e si sviluppa all’interno di specifici contesti sociali e culturali. Sotto il profilo epistemologico, la probabilità sostituisce via via sempre più la certezza e ogni conoscenza si configura come approssimazione rivedibile. Questo approccio contrasta con le esigenze della comunicazione pubblica, che predilige semplificazioni e slogan. La statistica, invece, lavora con astrazioni raffinate che, pur senza corrispettivi empirici diretti, consentono una più profonda comprensione della realtà. In quanto “arte critica”, essa non conferma categorie analitiche precostituite, ma anzi tende a metterle in crisi, promuovendo un’attitudine nei confronti del dato che mantiene aperto quanto più possibile il ventaglio del giudizio e la flessibilità interpretativa. Così intesa, la statistica non solo descrive il mondo, ma partecipa attivamente alla costruzione di visioni più ricche, dinamiche della realtà.
The Interplay between Bayesian Inference and Conformal Prediction
October 17, 2025, 12:00
Conformal prediction has emerged as a cutting-edge methodology in statistics and machine learning, providing prediction intervals with finite-sample frequentist coverage guarantees. Yet, its interplay with Bayesian statistics – often criticised for lacking frequentist guarantees – remains underexplored. Recent work has suggested that conformal prediction can serve to “calibrate” Bayesian procedures, thereby imparting frequentist validity and motivating deeper investigation into frequentist–Bayesian hybrids. We further argue that Bayesian inference has the potential to enhance conformal prediction, for instance, through more informative intervals. Thus, in a spirit of modern statistics based less on division and more on synthesis, the two paradigms may be viewed as complementary, jointly striving for a principled balance between validity and efficiency. In this talk, I will outline paths toward bridging this gap. After surveying existing ideas, a formalization of the Bayesian conformal inference framework will be provided, in both its full and split forms. Emphasis will be given to the challenging aspect of computational complexity, with potential solutions. Finally, the advantages of Bayesian conformal inference will be discussed in small-area estimation, a paradigmatic example for this hybrid perspective. Based on joint work with Brunero Liseo.
Plug-in estimation for parametric and penalised multi-state Markov models using ordinary differential equations
October 3, 2025, 12:00
Existing multi-state models are restricted either by the 'plug-in' parameters that can be estimate, or the dependence on the bootstrap for variance estimation. Our objective is to develop an efficient algorithm and implementation for 'plug-in' maximum likelihood estimation for parametric and smooth penalised Markov multi-state models. For methods, we restrict our attention to smooth parametric and penalised transition intensities for multi-state Markov models. We propose a new algorithm that uses a system of ordinary differential equations to calculate the parameters and their gradients, with standard errors calculated using the delta method. The algorithm supports 'plug-in' parameter estimation for state occupancy probabilities, transition probabilities, length of stay, relative survival, screening sensitivity, utilities, costs, net monetary benefit and their linear combinations. We provide an implementation in R that allows for a wide range of parametric and penalised survival models. Using simulations, we demonstrate good coverage for a range of transition intensity models. We apply these methods to an earlier multi-state analysis of the Rotterdam Breast Cancer Data and extend the analysis to include regression standardisation. In conclusion, we have provided a broad framework for 'plug-in' parameter estimation for Markov multi-state models with smooth transition intensities. These methods have applications to a range of disciplines, including epidemiology and health economics.
An Overview of Robust Regression Mixture Models Emphasizing Symmetric α-Stable Distributions
September 23, 2025, 12:00
The typical method for estimating mixture of regression models relies on the assumption that error components are normally distributed. This assumption makes them highly vulnerable to outliers or data with heavy-tailed errors. This lecture will review some robust alternatives for mixture of regression models. In particular, we will focus on a new robust model introduced by Zarei, which extends the mixture of symmetric α-stable (SαS) distributions to the regression setting. The SαS distribution is a heavy-tailed generalization of the normal distribution, where an additional parameter, α, controls the heaviness of the tails. A unique characteristic of the SαS distribution is that its variance diverges to infinity when α < 2. This property makes the model exceptionally robust against extreme outliers compared to other heavy-tailed distributions, like the Student's t-distribution. The model's parameters, except for α, are estimated using a standard Expectation-Maximization (EM) algorithm. The parameter α is estimated separately via a stochastic EM algorithm that utilizes a rejection sampling method. We will illustrate and compare this new model with existing mixture regression models using both simulated and real-world datasets.
Bi-clustering multivariate categorical data via extended mixtures of latent trait analyzers
May 30, 2025, 12:00
Multivariate categorical outcomes are common in real-world applications but, frequently, their high dimensionality makes both the analysis and the interpretation particularly challenging. In this context, model-based clustering offers a powerful approach to data reduction and structure discovery. We build upon the Mixture of Latent Trait Analyzers (MLTA) framework to propose a model that enables simultaneous clustering of both rows and columns of the data matrix. Specifically, rows are grouped into components using a finite mixture specification. Within each component, variables are segmented based on a flexible yet parsimonious specification of the linear predictor. To capture residual dependence among observations, we retain the use of a multidimensional latent trait, consistent with the original MLTA formulation. Additionally, the model accommodates the influence of individual-specific covariates on the clustering process via a concomitant variable framework. Parameter estimation is carried out using maximum likelihood, implemented through an extended Expectation-Maximization (EM) algorithm. Since the likelihood involves integrals without closed-form solutions, we apply a Gauss-Hermite quadrature for numerical approximation. A comprehensive simulation study evaluates the model’s ability to accurately recover both clustering structure and parameter values, demonstrating strong performance. Finally, we apply the proposed method to an original dataset on pediatric patients with suspected appendicitis, aiming to identify patient subgroups characterized by distinct patterns of clinical conditions.
On the Estimation of Climate Normals and Anomalies
May 23, 2025, 12:00
The quantification of the interannual component of variability in climatological time series is essential for the assessment and prediction of the El Niño - Southern Oscillation phenomenon. This is achieved by estimating the deviation of a climate variable (e.g., temperature, pressure, precipitation, or wind strength) from its normal conditions, defined by its baseline level and seasonal patterns. Climate normals are currently estimated by simple arithmetic averages calculated over the most recent 30-year period ending in a year divisible by 10. The suitability of the standard methodology has been questioned in the context of a changing climate, characterized by nonstationary conditions. The literature has focused on the choice of the bandwidth and the ability to account for trends induced by climate change. The paper contributes to the literature by proposing a regularized real time filter based on local trigonometric regression, optimizing the estimation biasvariance trade-off in the presence of climate change, and by introducing a class of seasonal kernels enhancing the localization of the estimates of climate normals. Application to sea surface temperature series in the Niño 3.4 region and zonal and trade winds strength in the equatorial and tropical Pacific region, illustrates the relevance of our proposal. Joint work with Alessandro Giovannelli, Università dell’Aquila.
A Hierarchical Model for Comparing Spectral Patterns in Lemur Vocalization
May 16, 2025, 12:00
In this talk, I will present a hierarchical model designed to analyze the spectrograms of animal vocalizations, with a focus on grunt calls from different lemur species. The primary goal is to uncover a latent spectral shape that characterizes each species and allows us to quantify dissimilarities between them. A key challenge lies in aligning calls of varying durations and temporal dynamics. To tackle this, we incorporate a synchronization function to manage non-stationary temporal features and adopt a circular representation of time to handle artifacts caused by the discretization of analog signals. Given the high dimensionality of spectrogram data, we use a Nearest Neighbor Gaussian Process for efficient computation and sample from the posterior distribution using MCMC. The model is applied to recordings from eight lemur species. For each species, we identify a representative vocal pattern and use a simple distance metric to compare them. Predictive performance is assessed via crossvalidation, and we also explore some special cases that highlight the model’s flexibility.
Mixture-based clustering for ordinal responses
April 4, 2025, 12:00
Existing methods can perform likelihood-based clustering on a multivariate data matrix of ordinal data, using finite mixtures to cluster the rows (observations) of the matrix. These models can incorporate the main effects of individual rows and columns, as well as cluster effects, to model the matrix of responses. However, many real-world applications also include available covariates, which can provide insights into the main characteristics of the clusters. In our research, we have extended the mixture-based models to include covariates directly, to allow the clustering structures to be determined both by the individuals' similar patterns of responses and the effects of the covariates on the individuals' responses. We focus on clustering the rows of the data matrix, using the proportional odds cumulative logit model for ordinal data. We fit the models using the Expectation-Maximization algorithm and assess performance through a comprehensive simulation study. We also illustrate an application of the models.
High-Dimensional Covariance Estimation via Sparse Pairwise Likelihood
March 28, 2025, 12:00
Pairwise likelihood offers a practical approximation to the full likelihood function, enabling efficient inference for high-dimensional covariance models by combining marginal bivariate likelihoods. This approach simplifies complex dependencies and retains optimal statistical efficiency in certain models, such as the multivariate normal distribution, where pairwise and full likelihoods are maximized by the same parameter values. We propose a novel method for estimating sparse high-dimensional covariance matrices by maximizing a truncated pairwise likelihood function, which includes only terms corresponding to nonzero covariance elements. Truncation is achieved by minimizing the L2-distance between pairwise and full likelihood scores, coupled with an L1-penalty to exclude uninformative terms. Unlike traditional regularization techniques, our method focuses on selecting entire pairwise likelihood objects rather than shrinking individual parameters, preserving the unbiasedness of the estimating equations. Theoretical analysis demonstrates that the resulting estimator is consistent and achieves the same efficiency as the oracle maximum likelihood estimator, which assumes knowledge of the nonzero covariance structure, even as the dimensionality grows exponentially. Numerical experiments confirm the effectiveness of the proposed approach.
Extensions of Item Response Theory models for analysing complex phenomena
March 21, 2025, 12:00
Item Response Theory (IRT) provides a robust framework for analysing phenomena that cannot be directly observed. In such cases, inference relies on observable behaviours, such as responses to questionnaire items or survey indicators related to the phenomenon of interest. This talk will introduce key IRT concepts, with a particular focus on extensions suited for analysing complex phenomena. Specifically, the benefits of modelling the latent trait with a discrete distribution will be discussed, which, unlike traditional IRT models based on a continuous trait, offer a more flexible representation of heterogeneous populations. Moreover, the talk will cover the concept of multidimensionality, which allows for the representation of multiple latent traits through a multivariate latent variable approach. This is particularly useful when analysing constructs that are inherently composed of several interrelated dimensions. The presentation will conclude with two real-world applications: assessing the risk profiles of healthcare contracting authorities in public procurement management and estimating educational poverty levels in Italy by integrating small area estimation with multidimensional IRT.
Environmental risk assessment via concomitant-variable multivariate penalized hidden semi-Markov models with autoregression
March 7, 2025, 14:00
Environmental risk assessment often requires modeling complex temporal processes influenced by multiple variables and characterized by environmental condition shifts. This work introduces a novel methodological framework based on concomitant-variable multivariate penalized hidden semi-Markov models (CV-MPHSMM) with autoregression to capture such dynamics. The proposed model extends traditional hidden semi-Markov models by integrating concomitant variables to account for external environmental factors influencing state transitions and sojourns, and by incorporating penalization techniques to enhance model interpretability and prevent overfitting in high-dimensional settings. Autoregressive components are included to model temporal dependencies within and between observed multivariate time series. Analytical expressions for multivariate risk measures are obtained under the CV-MPHSMM. The framework is applied to pollution, demonstrating its capacity to identify latent states, quantify transition probabilities, and detect environmental condition shifts. Simulation studies validate the robustness and flexibility of the proposed model in handling complex scenarios, while case studies highlight its practical utility in informing risk management strategies. The findings underscore the potential of CV-MPHSMMs with autoregression as a powerful tool for advancing environmental risk assessment and decision-making under uncertainty.
Linkages between AIR POLLUTION and CLIMATE CHANGE: MODELS and MEASUREMENTS
January 31, 2025, 12:00
Air pollution and climate change are closely linked: the chemical species that lead to a degradation in air quality are normally co-emitted with greenhouse gases. Thus, changes in one inevitably cause changes in the other. Air pollution and climate change are both threats to global population health and require a response that involves intersectoral policy and action. What are the tools commonly used to tackle these issues? Since the 90’s, ENEA has been established long term observatories of the main climatic parameters in “hot spot” regions like the Mediterranean and Antarctica. The Station for Climate Observations on the island of Lampedusa, an integrated research facility in the Mediterranean, has been collecting more than 25 years of data for the study of climate change. In the same years, ENEA has developed a chemical transformation model to study air pollution producing different simulations both for present and future scenarios (till the year 2030-2050), than for 3-days forecast over Italy (4km) and Europe (10km). During the seminar an overview of the models developed, the measures collected in these remote regions and the activities currently carried out will be presented, discussing both long time series than specific studies, focusing on the main challenges and uncertainties the scientific community is facing.
Language development in pre-school children: study designs and methods of analysis (In Italian)
January 24, 2025, 12:00

2024


Does wind shape vegetation stripes? A copula-based mixture model for axial and circular data
December 6, 2024, 12:00
Motivated by a case study of vegetation patterns, we introduce a mixture model with concomitant variables to examine the association between the orientation of vegetation stripes and wind direction. The proposal uses a novel copula-based bivariate distribution for mixed axial and circular observations. It provides a parsimonious and computationally tractable approach to examining the dependence of two environmental variables observed in a complex manifold.
Dealing with varying dimension-wise tail weights in normal scale mixtures
November 29, 2024, 12:00
We present the family of dimension-wise scaled normal mixtures (DSNMs). This family is designed to model the joint distribution of a d-variate random variable with real-valued components, extending the multivariate normal (MN) distribution in two directions. First, DSNMs have a more general type of symmetry with respect to the elliptical one. Secondly, the univariate marginals have similar heavy-tailed normal scale mixture distributions with (possibly) different tailedness parameters, allowing for dimension-specific excess kurtosis. This flexibility is achieved within an MN scale mixture framework by incorporating a d-variate mixing random variable with independent and similar components acting separately for each dimension. We explore several properties of DSNMs, including their joint density function, hierarchical and stochastic representations, relationships with other symmetric distributions, symmetry type, marginal distributions, no correlation implying independence, and moments of practical interest. For illustrative purposes, we describe two members of the DSNM family obtained in the case of components of the mixing random vector being either uniform or shifted exponential; these are examples of mixing distributions that guarantee a closed-form expression for the joint density of the DSNM. For the two DSNMs analyzed in detail, we introduce parsimony by allowing the d tailedness parameters to be tied across dimensions and describe algorithms, based on the expectation–maximization (EM) principle, to estimate the parameters by maximum likelihood. Finally, we apply DSNMs to real financial and biometrical data, demonstrating their advantages over existing symmetric heavy-tailed distributions in the literature.
SEMI-PARAMETRIC ESTIMATION OF A SPATIO-TEMPORAL HAWKES PROCESS FOR MODELLING CAR ACCIDENT
November 22, 2024, 12:00
We propose a semi-parametric spatio-temporal Hawkes process to model the occurrence of car accidents over general domains. The overall intensity is split into the sum of a background component capturing the spatio-temporally varying intensity and an excitation component accounting for the possible triggering effect between events. The spatial background is estimated and evaluated on an approximation to the road network, allowing the derivation of accurate risk maps of road accidents. We ensure that the spatio-temporal excitation preserves an isotropic behavior and we generalize it to account for the effect of covariates. The estimation is pursued maximizing the expected complete data log-likelihood using a tailored version of the stochastic reconstruction algorithm that adopts ad-hoc boundary correction strategies. Two applications analyse car accidents occurred on the London M25 Orbital in 2018 and on the Rome road network in years 2019, 2020, and 2021. Results highlight that car accidents of different types might exhibit varying degrees of excitation, ranging from no triggering to an 8% chance of triggering one further event.
FUNCTIONAL FACTOR ANALYSIS OF BRAIN CONNECTIVITY
November 15, 20234, 12:00
In this presentation we explore the development and application of factor analysis with functional data. Functional data analysis, which involves data repeatedly measured over some domain (such as space or time), often makes use of dimension reduction techniques such as principal component analysis, but factor analysis has thus far been relatively under explored in this area. Motivated by applications to fMRI which require interpretable dimension reduction tools, we build a factor analysis model suitable for such data which provides interpretable components that scientists can make use of in their downstream analyses.
AN SSA-BASED PROCEDURE FOR EXPLORING STRUCTURAL CHANGES IN TIME SERIES
November 12, 2024, 12:00
Some procedures based on Singular Spectral Analysis focus on detecting potential structural changes in a time series by comparing a single decomposition method applied to two trajectory matrices (base and test). We propose a new approach that compares two different decomposition methods, both using the NIPALS algorithm, applied to the same trajectory matrix of the time series. Examples are provided to illustrate this new structural change detection procedure. (a joint work with Alberto Silva) Some procedures based on Singular Spectral Analysis focus on detecting potential structural changes in a time series by comparing a single decomposition method applied to two trajectory matrices (base and test). We propose a new approach that compares two different decomposition methods, both using the NIPALS algorithm, applied to the same trajectory matrix of the time series. Examples are provided to illustrate this new structural change detection procedure.
How to address bias and confounding when biological sex is the exposure
October 25, 2024, 12:00
Background: Global child mortality rates vary by sex, with higher mortality rates generally found in males. However, previous research has shown that this ratio is reversed in infants admitted to Paediatric Intensive Care Units (PICU). I aimed to determine whether female sex is causally linked to higher mortality in PICU. Methods: I created a longitudinal linked dataset that could be used to evaluate whether sex is causally related to mortality in PICU using routine data on >100,000 children admitted to PICU. I compared a number of estimation methods, namely (i) gcomputation, (ii) propensity score-based singly and doubly robust methods, and (iii) targeted learning aided by machine learning, to determine whether the observed sexratio reversal in PICU mortality is supported by the data. Results: Female biological sex increased the mortality rate in PICU by up to 0.26% (95% CI -0.05%, 0.57%). In the multiply imputed dataset this estimate was 0.35% (95% CI 0.09%, 0.61%). The reversal in mortality rates in PICU was not explained by collider bias. The collider bias was driving the naïve estimate towards the null. Conclusion: Female biological sex is linked to higher mortality in PICU. Mechanistic reasons underlying this causal relationship are still unknown.
Online Multivariate Changepoint Detection: Leveraging Links With Computational Geometry
October 22, 2024, 12:00
The increasing volume of data streams poses significant computational challenges for detecting changepoints online. Likelihood-based methods are effective, but a naive sequential implementation becomes impractical online due to high computational costs. We develop an online algorithm that exactly calculates the likelihood ratio test for a single changepoint in pdimensional data streams by leveraging fascinating connections with computational geometry. This connection straightforwardly allows us to recover sparse likelihood ratio statistics exactly: that is assuming only a subset of the dimensions are changing. Our algorithm is straightforward, fast, and apparently quasi-linear. A dyadic variant of our algorithm is provably quasi-linear, being Op(nlog(n) p+1) for n data points and p less than 3, but slower in practice. These algorithms are computationally impractical when p is larger than 5, and we provide an approximate algorithm suitable for such p which is Op(np̃log(n) p̃+1), for some user-specified p̃≤ 5. We derive some statistical guarantees for the proposed procedures in the Gaussian case, and confirm the good computational and statistical performance, and usefulness, of the algorithms on both empirical data and on NBA data.
Depths and Local Depths. A journey in statistical depths
October 4, 2024, 12:00
Statistical data depth is a growing area in non-parametric statistics, originally developed for the analysis of multidimensional data but useful in other frameworks such as e.g. spherical and functional data. The main applications are a center-outward ordering of the observations, location and scale estimation, classification, clustering and some graphical tools. Statistical local depth functions are a generalization of statistical depth functions and they are used for describing local geometric features and mode(s) in multivariate distributions. In this seminar, after an introduction on statistical data depths, we illustrate some analytical and statistical properties of the local depths. We show how these functions are a bridge between density functions and depths, we illustrate their theoretical properties and we discuss some applications.
A time-heterogeneous rectangular latent Markov model with covariates to measure dynamics and correlates of student's learning abilities data
June 3, 2024, 12:00
Accurate and up-to-date assessments of students' abilities are essential for personalised learning. These assessments allow instructors to adjust class content to match different skill levels, and help students gain awareness of their learning paths. While technology-based learning environments have eased data collection, arguably they have brought important challenges for analysis, mainly relating to data complexity. This work introduces a novel fully time-heterogeneous rectangular latent Markov specification, tailored to complex longitudinal data of this kind. The proposed toolkit incorporates measurement model heterogeneity, allowing for possibly as many different measurement models as the number of distinct measurement occasions. The structural model, in which we include predictors of initial and transition probabilities, is consequently specified, and informative dropout is modelled explicitly and jointly with its potential correlates. However, the resulting model is overly complex to estimate with standard simultaneous procedures. We address the estimation problem by designing a bias-adjusted three-step estimator, which separates the estimation of the measurement models from the structural model fit. Our primary empirical aim is to analyse the abilities and progression in learning statistical topics over time of a concrete cohort of students, while accounting for their individual characteristics. Results from an extensive simulation study substantiate our empirical findings.
Soil consumption and organized crime
May 24, 2024, 12:00
Soil is a non-renewable natural resource providing ecosystem services essential for life. Soil consumption is the increase in artificial land cover through anthropogenic activities. While standard economic variables (population and GDP growth) appear to have limited predictive power for soil consumption, the recent literature on the mutually beneficial relationship between criminal organizations (“mafias”) and local politicians/administrators suggest a role for the presence and strength of mafias at the local level. We contribute to the literature by providing direct evidence of the link between soil consumption and mafia strength in the Southern Italian region of Apulia using a rich dataset at the fine municipality level that we created by merging information from a variety of sources. We show that alternative measures of the local strength of organized crime help improve substantially our predictions of soil consumption, both total soil consumption and soil consumption in protected areas. Under a plausible instrumental variable assumption, we also provide a quantitative assessment of the causal effect of the local strength of organized crime on soil consumption.
On the complexity for functional data
May 17, 2024, 12:00
The complexity of a stochastic process may be associated with the notion of degrees of freedom and, in some situations, may coincide with the concept of dimensionality. In this talk, complexity is studied through the use of the small-ball probability of the process. Specifically, we assume that this probability factorizes in two terms: one dependent only on the center of the ball, the other dependent only on the radius. The second term, which includes the information about the complexity, will be studied statistically: estimation, asymptotic behavior, and practical performance in a simulated and real environment.
Introduzione all’Analisi dei Dati Spaziali
May 15, 2024, 10:00-14:00
Seminari del Dottorato - Scuola di Scienze Statistiche - Curriculum Demografia (the seminar will be held in Italian)
Issues and challenges in making inference from non-probability samples
May 3, 2024, 12:00
In the recent decade the relevance of the non-probability sampling in surveys is considerably increased because of the availability of alternative data sources such as Big Data and web surveys. The major concern about non-probability samples is that the unknown selection process is frequently selective, so that they often fail to represent the target population properly and hence result in highly biased estimators. In this work two approaches for dealing with selection bias when the selection process is nonignorable are discussed. The first one, based on the empirical likelihood, does not require parametric specification of the population model but the probability of being in the non-probability sample needed to be modeled. Auxiliary information known for the population or estimable from a probability sample can be incorporated as calibration constraints, thus enhancing the precision of the estimators. The second one introduces the concept of uncertainty on data generating model resulting from the lack of knowledge of the sampling design acting in the nonprobability sample. First of all, when extra-sample information is available, the class of plausible distributions for the variable of interest is defined. Next, a plausible estimate for such distribution is constructed and its accuracy is evaluated.
External information borrowing in clinical trial hypothesis testing with controlled TIErate inflation
April 5, 2024, 12:00
When designing a novel clinical trial, external information about the control and/or treatment arm effect is typically available. Borrowing of such external information is often desired in order to improve the trial’s efficiency, and can be of crucial importance in situations where the sample size that can realistically be recruited is limited, as, e.g., pediatric or rare disease trials. The Bayesian approach allows borrowing of such external information through the adoption of informative prior distributions. An issue associated with the incorporation of external information is that external and current information may systematically differ. However, such inconsistency may not be predictable or quantifiable a priori. Robust prior choices are typically proposed to avoid extreme worsening of operating characteristics in such situations. In this talk, we will focus on frequentist type I error rate and power. We will in particular consider how type I error rate is affected by incorporation of external information, and present a novel approach which allows a principled and controlled inflation. Both one and two-arm clinical trial designs will be considered.
Does measurement error distort country differences in temporary employment? differences in temporary employment? A study on Italy and the Netherlands A study on Italy and the Netherlands using a multi- using a multi-group hidden Markov model group hidden Markov model
March 22, 2024, 12:00
This paper investigates the effect of measurement error on two key labour market indicators: the distribution of temporary employment as well as the transition rate in and out of temporary employment over time in Italy and the Netherlands. In this way, we study whether the cross-country differences in these indicators (coming from the different institutions in the two labour markets) pertain when we correct for measurement error in socioeconomic data. The comparative analysis of the Italian and Dutch labour markets is carried out for the time period 2017-2019 using linked employment data from the Labour Force Survey and the Employment Register of the two countries. For this purpose, we use a multiple-group Hidden Markov Model with two indicators for the employment contract type that accounts for both random and systematic measurement error. The results indicate that measurement error severely biases our view on mobility from temporary to permanent employment in the two countries but also distorts the picture of cross-country differences in the phenomenon of interest.
Finite Population Survey Sampling: An unapologetic Bayesian Perspective
March 1, 2024, 18:00
In this talk I will offer some perspectives on Bayesian inference for finite population quantities when the units in the population are assumed to exhibit complex dependencies. Beginning with an overview of Bayesian hierarchical models, including some that yield design-based Horvitz-Thompson estimators, the talk proceeds to introduce dependence in finite populations and sets out inferential frameworks for ignorable and nonignorable responses. Multivariate dependencies using graphical models and spatial processes are discussed and some salient features of two recent analyses for spatially oriented finite populations are presented.

2023


Ten years of mobile phone big data statistical analyses
December 1, 2023, 12:00
In the era of big data, monitoring and forecasting people crowding and mobility is a relevant aspect for urban policies, and smart cities use signals from mobile phone networks to support the optimization of urban systems and flows. Mobile phone data can be used for various purposes, as they come in different types: in this talk, applications are presented for social and cultural events monitoring, mobility flows and flooding risk analysis. Special attention is devoted to the statistical methods useful for these analyses, based on spatio-temporal data: briefly discussed are results obtained using the Histogram of Oriented Gradients approach for image reduction, the Functional Data Clustering of time series and the VARX model with Harmonic Dynamic Regression.
From Rome to London: Breaking the Silos of Disciplines. A Conversation around Cardiometabolic Risks Factors Modelling and the Use of Wastewater in Public Health
November 27, 2023, 10:00-12:00

Circular local likelihood regression
November 24, 2023, 12:00
In this talk, we will present a general framework for estimating regression models with circular covariates and a general response. We will start with an overview on nonparametric regression models with circular covariate, revising the main ideas and motivating the need of a more general method. Our goal is to estimate (nonparametrically) a conditional characteristic by maximizing the circular local likelihood. The proposed estimator is shown to be asymptotically normal. The problem of selecting the smoothing parameter is also addressed, as well as bias and variance computation. The finite sample performance of the estimation method in practice is studied through an extensive simulation study, where we cover the cases of Gaussian, Bernoulli, Poisson and Gamma distributed responses. The generality of our approach is illustrated with several real-data examples from different fields. In particular, we will focus on an example of neural response in macaques. This is a joint work with M. Alonso-Pena and I. Gijbels and corresponds to two published papers in Biometrics (2023) and Journal of the American Statistical Association (2023).
Surveying sensitive topics with indirect questioning techniques: methods and real applications
November 17, 2023, 12:00
In many fields of the applied research, mostly in sociological, economics, demographic, ecological and medical studies, the investigator very often has to gather information concerning highly personal, sensitive, stigmatizing, and perhaps incriminating issues such as drug addiction, domestic violence, racial prejudice, illegal income, noncompliance with laws and regulations. Doing research on sensitive themes by traditional direct questioning survey modes is not an easy matter since it is likely to meet with two sources of errors: nonresponse and untruthful answers. These errors can seriously flaw the quality of the data and, thus, jeopardize the usefulness of the collected information for subsequent analyses including inference on unknown characteristics of the population under study. Although the errors cannot be totally avoided, they may be mitigated by increasing respondent cooperation through a nonstandard data-collection approach based on indirect questioning techniques (IQTs). The talk aims to introduce some issues related to privacy protection when sensitive topics are surveyed, give a general idea of the approach, and illustrate how some IQTs have been used to collect data and obtain prevalence estimates in a number of real studies about illegal immigrants, abortion, drug use, cannabis legalization, sexual behaviours and Covid-19 health behaviours.
Specification tools for time fixed effects in country panels
November 13, 2023, 15:00
This paper proposes specification tools for time fixed-effects in country panels that complement summary and graphical representations of the data. They cover the standard two-way fixed-effect model, as well as other more general d-way fixed effect specifications with d ≥ 3. The tools are based on the observable characteristics of univariate time series of contrasts implied by a given specification; they use flagging rules based on graphical or statistical analysis. Evidence on which contrasts do not contain time fixed-effects can be harvested by algorithms; this paper discusses two examples of such algorithms. Implications for the specification of Differences in Differences estimation are discussed, and results are illustrated using a country panel of prices of mobile telecommunication services.
A parsimonious family of mixtures of multivariate Poisson log-normal factor analyzers for clustering count data
November 3, 2023, 15:00
Multivariate count data are commonly encountered in bioinformatics. Although the Poisson distribution seems a natural fit for these count data, its multivariate extension is computationally expensive. Recently, mixtures of multivariate Poisson lognormal (MPLN) models have been used to efficiently analyze these multivariate count measurements. In the MPLN model, the counts, conditional on the latent variable, are modelled using a Poisson distribution, and the latent variable comes from a multivariate Gaussian distribution. Due to this hierarchical structure, the MPLN model can account for over-dispersion as opposed to the traditional Poisson distribution and allows for correlation between the variables. The mixture of multivariate Poisson-log normal distributions for high dimensional data is extended by incorporating a factor analyzer structure in the latent space. A family of parsimonious mixtures of multivariate Poisson lognormal distributions are proposed by decomposing the covariance matrix and imposing constraints on these decompositions. The performance of the model is demonstrated using simulated and real datasets.
Testing for treatment effect in multitreatment case
October 27, 2023, 12:00
In the present seminar, the problem of testing for the presence/absence of a treatment effect is discussed. A new test-statistic, essentially based on the same principles as the classical Kruskal-Wallis test, is introduced, and its theoretical properties are studied. The good behaviour of the proposed test in terms of both significance level and power, with respect to other commonly used test procedures, is shown through a simulation study. Test-statistics for stochastic dominance problems are also studied.
Hidden Markov Models for error corrected statistics
October 20, 2023, 12:00
Policy making is based on official statistics that may still include measurement error (ME). This ME is typically the result of administrative delays in registration, differenced in conceptual definitions or processing errors. ME can lead to a distorted view of the number of people in groups of interest. Threatening the integrity of official statistics and therefore also the effectiveness of policies based on them. HMMs are statistical models that help approximate categorical variables that are incorrectly measured with observed data. First, HMMs use two observed measures of the same statistic from different sources to approximate the error-corrected number of individuals at every moment in time. The two observed measures help to triangulate “true” information and reduce the effect of ME that exists in each of them. Second, HMMs estimate the transition rate from one state (e.g., receiving social assistance benefits) to other states (e.g. employment) and vice versa using the error-corrected measure that was approximated before. To do so, HMMs require several (at least three) observations in different time points per individual, which are available in our case in both registers. This way HMM present a measurement model (error correction), and structural model (representing the relations and changes over time).
The resilience of complex networks: methods and applications
October, 6 2023
The analysis of the resilience of a network is of key relevance in many contexts of applied science, for its natural connections with the assessment of the stability of an overall system. In this talk I will present some methodological criteria for building suitable resilience measures, along with some applicative instances. I will also provide some remarks on avenues of future research, by including also a discussion on the possible connections between complex networks and reliability theory.
Time series segmentation by non-homogeneous hidden semi-Markov models
May, 26, 2023
Motivated by classification issues in environmental studies, a class of hidden semi-Markov models is introduced to segment multivariate time series according to a finite number of latent regimes. The observed data are modelled by a mixture of multivariate densities, whose parameters evolve according to a latent multinomial process. The multinomial process is modelled as a semi-Markov chain where the time spent in a state and the chances of a regime- switching event are separately modeled by a battery of regression models that depend on time- varying covariates. Maximum likelihood parameter estimation is carried out by integrating an EM algorithm with a suitable data augmentation. While the proposal extends previous approaches that rely on mixtures models and hidden Markov models, it keeps a parsimonious structure that facilitates results interpretation. It is illustrated on a case study of a bivariate time series of wind and wave directions, observed by a buoy in the Adriatic sea.
ABCC : Approximate Bayesian Conditional Copulae
April 19 2023
Copula models are flexible tools to represent complex structures of dependence for multivariate random variables. According to Sklar's theorem any d-dimensional absolutely continuous density can be uniquely represented as the product of the marginal distributions and a copula function that captures the dependence structure among the vector components. In real data applications, the interest of the analyses often lies on specific functionals of the dependence, which quantify aspects of it in a few numerical values. A broad literature exists on such functionals, however, extensions to include covariates are still limited. This is mainly due to the lack of unbiased estimators of the copula function, especially when one does not have enough information to select the copula model. Recent advances in computational methodologies and algorithms have allowed inference in the presence of complicated likelihood functions, especially in the Bayesian approach, whose methods, despite being computationally intensive, allow us to better evaluate the uncertainty of the estimates. In this work, we present two Bayesian methods to approximate the posterior distribution of functionals of the dependence, using nonparametric models which avoid the selection of the copula function. These methods are compared in simulation studies and in a realistic application in astrophysics. Joint work with Clara Grazian and Luciana Dalla Valle.
Convex clustering of mixed numerical and categorical data
April, 28 2023
Clustering analysis is an unsupervised learning technique widely used for information extraction. Current clustering algorithms often face instabilities due to the non-convex nature of their objective function. The class of convex clustering methods does not suffer from such instabilities and finds a global optimum for the clustering objective. Whereas convex clustering has previously been established for single-type data, real-life data sets usually comprise both numerical and categorical, or mixed, data. Therefore, we introduce the mixed data convex clustering (MIDACC) framework. We implement this framework by developing a dedicated subgradient descent algorithm. Through numerical experiments, we show that, in contrast to baseline methods, MIDACC achieves near-perfect recovery of both spherical and non-spherical clusters, is able to capture information from mixed data while distinguishing signal from noise, and has the ability to recover the true number of clusters present in the data. Furthermore, MIDACC outperforms all baseline methods on a real-life data set.
The Multivariate Heterogeneous Preference Estimation for Switching Multiple Price List
April 21, 2023
The Multivariate Heterogeneous Preference Estimator for Switching Multiple Price List Switching Multiple Price List Anna Conte Dipartimento di Scienze Statistiche, Sapienza Università di Roma The Multiple Price List (MPL) and Switching Multiple Price List (sMPL) provide a useful framework for estimating preference parameters, most usually risk aversion, from a sample of experimental subjects or survey respondents. In this paper, we consider designs in which more than one sMPL is presented to each subject, allowing more than one preference parameter to be estimated simultanously, and we propose a consistent estimator in this setting - the Multivariate Heterogeneous Preference (MHP) estimator. Focusing on the bivariate case of two sMPLs and two preference parameters, we demonstrate that non-standard econometric techniques, namely Monte Carlo integration with importance sampling, are required to implement the MHP estimator. Via a Monte Carlo exercise, we show that our estimator has good finite-sample properties. Finally, we apply the MHP estimator to a real data set and compare the estimates to those obtained using an inconsistent estimator applied in previous studies.
Directional distribution depth function and its application to classification
April 14, 2023
Statistical depth functions are introduced as a way to provide a center-outward ordering of the sample points in multidimensional space, which can be used for outlier detection, classification, and other exploratory tools. In this work we propose a novel definition of depth function for multivariate data using random directions, which preserves the Mahalanobis distance of the points in the original space. More specifically, the proposed depth function is the expectation of all depths along the potentially infinite random directions, which, in turn, are functions of the point percentiles estimated via parametric or nonparametric models. The proposed method is evaluated through simulated experiments and real data applications, and is shown to be effective in supervised classification problems.
Functional estimation of anisotropic covariance and autocovariance operators on the sphere
March 31, 2023
In this talk we present nonparametric estimators for the second-order central moments of possibly anisotropic spherical random fields, within a functional data analysis context. We consider a measurement framework where each random field among an identically distributed collection of spherical random fields is sampled at a few random directions, possibly subject to measurement error. The collection of random fields could be i.i.d. or serially dependent. Though similar setups have already been explored for random functions defined on the unit interval, the nonparametric estimators proposed in the literature often rely on local polynomials, which do not readily extend to the (product) spherical setting. We therefore formulate our estimation procedure as a variational problem involving a generalized Tikhonov regularization term. Using the machinery of reproducing kernel Hilbert spaces, we establish representer theorems that fully characterize the form of our estimators. We determine their uniform rates of convergence as the number of random fields diverges, both for the dense (increasing number of spatial samples) and sparse (bounded number of spatial samples) regimes. A simulation study and a preliminary exploration of a real dataset of ocean temperatures will be also discussed. Joint work with Julien Fageot, Matthieu Simeoni and Victor M. Panaretos.
A long history of model based clustering based on trimming and constraints
March 24, 2023, 12:00
Model based clustering plays a major role in data analysis. Our interest focuses on approaches related to maximum likelihood estimation via EM/CEM algorithms. However, it is very common that input-datasets contain observations belonging to contaminating sources, out of the assumed family of distributions in the chosen model. It is well known that this contamination in the sample is able to break likelihood based estimators. Methodology based on the joint application of trimming and constraints, under the label TCLUST, has been developed for robustifying model based clustering proposals. Trimming tries to eliminate contaminating observations, however in order to achieve robust proposals in clustering, it is also needed to apply constraints to control the relative size of clusters' variability. There are TCLUST procedures available for estimating mixtures in different settings: linear models, factor analyzers and functional data among others. Statistical properties of TCLUST procedures, including consistency and a non negligible breakdown point are available. TCLUST's constraints have evolved in the last few years, providing an improved flexibility, in order to capture the patterns in the covariance matrix decomposition included in the classical parsimonious family of Celeux and Govaert. An important open issue in TCLUST procedures is related to their input parameters: the number of clusters, the level of trimming and the strength of the constraints. Exploratory tools and automatized procedures for assisting the users in choosing these input parameters have been developed. TCLUST procedures are available in CRAN ('tclust' package) and in MATLAB ('FSDA' toolbox).
Causal Regularization
March, 17, 2023 (12 pm)
Causality is the holy grail of science, but humankind has struggled to operationalize it for millennia. In recent decades, a number of more successful ways of dealing with causality in practice, such as propensity score matching, the PC algorithm, and invariant causal prediction, have been introduced. However, approaches that use a graphical model formulation tend to struggle with computational complexity, whenever the system gets large. Finding the causal structure typically becomes a combinatorial-hard problem. In our causal inference approach, we build forth on ideas present in invariant causal prediction and the causal Dantzig and anchor regression, by replacing combinatorial optimization with a continuous optimization using a form of causal regularization. This makes our method applicable to large systems. Furthermore, our approach allows a precise formulation of the trade-off between in-sample and out-of-sample prediction error. In allegato la locandina con l'abstract e i riferimenti per partecipare in presenza o da remoto.
Central Quantile subspace and its applications
March 17, 2023 (4pm)
Quantile regression (QR) is becoming increasingly popular due to its relevance in many scientific investigations. There is a great amount of work about linear and nonlinear QR models. Specifically, nonparametric estimation of the conditional quantiles received particular attention, due to its model flexibility. However, nonparametric QR techniques are limited in the number of covariates. Dimension reduction offers a solution to this problem by considering low-dimensional smoothing without specifying any parametric or nonparametric regression relation. The existing dimension reduction techniques focus on the entire conditional distribution. We, on the other hand, turn our attention to dimension reduction techniques for conditional quantiles and introduce a new method for reducing the dimension of the predictor X. The performance of the methodology is demonstrated through simulation examples and data applications, especially to financial data. Finally, various extensions of the method are presented, such as nonlinear dimension reduction and the use of categorical predictors.
smoothEM: a new approach for smoothEM: a new approach for the simultaneous assessment of smooth patterns and spikes
10/03/2023
We consider functional data where an underlying smooth curve is composed not just with errors, but also with irregular spikes that (a) are themselves of interest, and (b) can negatively affect our ability to characterize the underlying curve. We propose an approach that, combining regularized spline smoothing and an ExpectationMaximization algorithm, allows one to both identify spikes and estimate the smooth component. Imposing some assumptions on the error distribution, we prove consistency of EM estimates. Next, we demonstrate the performance of our proposal on finite samples and its robustness to assumptions violations through simulations. Finally, we apply our proposal to data on the annual heatwaves index in the US and on weekly electricity consumption in Ireland. In both datasets, we are able to characterize underlying smooth trends and to pinpoint irregular/extreme behaviors. Work in collaboration with Huy Dang (Penn State University) and Francesca Chiaromonte (Penn State University and Sant’Anna School of Advanced Studies).
Spatio-Temporal Semantic -Temporal Semantic Partitions of the Land Surface through Deep Embeddings
03/03/2023
Temporal sequences of satellite images constitute a highly valuable and abundant resource to analyze a given region. However, the labeled data needed to train most machine learning models are scarce and difficult to obtain. In this context, we investigate a fully unsupervised methodology that, given a sequence of images, learns a semantic embedding and then, creates a partition of the ground according to its semantic properties and its evolution over time. We illustrate the methodology by conducting the semantic analysis of a sequence of satellite images of a region of Navarre (Spain). The proposed approach reveals a novel broad perspective of the land, where potentially large areas that share both a similar semantic and a similar temporal evolution are connected in a compact and well-structured manner. The results also show a close relationship between the allocation of the clusters in the geographic space and their allocation in the embedded spaces. The semantic analysis is completed by obtaining the representative sequence of tiles corresponding to each cluster, the linear interpolation between related areas, and a graph that shows the relationships between the clusters, providing a concise semantic summary of the whole region.

2022


Wrapping onto a torus: handling multivariate circular data in the presence of outliers
December 16, 2022, 12:00
Abstract: Multivariate circular data arise commonly in many different fields, including the analysis of wind directions, protein bioinformatics, animal movements, handwriting recognition, people orientation, cognitive and experimental psychology, human motor resonance, neuronal activity, robotics, astronomy, biology, physics, earth science and meteorology. Observations can be thought of as points on a p-dimensional torus, whose surface is obtained by revolving the unit circle in a p−dimensional manifold. The peculiarity of multivariate torus data is periodicity, that reflects in the boundedness of the sample space and often of the parametric space. The problem of modeling circular data has been tackled through suitable distributions, among which two of the most popular are the von Mises and the Wrapped Normal. Here, we focus on the family of unimodal and elliptically symmetric wrapped distributions with emphasis on the Wrapped Normal. Despite the boundedness of the support of circular variates, torus data are not immune to the occurrence of outliers, that is unexpected values, such as angles or directions, that do not share the main pattern of the bulk of the data. Then, a robust procedure to fit a wrapped distribution is presented. The proposed algorithm is characterized by the computation of data dependent weights aimed to down-weight anomalous values. We discuss and compare different approaches to obtain weights, with particular attention to the weighted likelihood methodology. A formal outliers detection rule is also suggested, that is based on classical robust distances evaluated over unwrapped data. In allegato la locandina con i riferimenti per partecipare al seminario e l'abstract.
Pseudo-populations resampling for finite populations under complex designs
December 2, 2022, 12:00
"Pseudo-populations resampling for finite populations under complex designs" Pier Luigi Conti Dipartimento di Scienze Statistiche Sapienza Università di Roma Abstract: The present talk is devoted to resampling for finite populations when the sampling design is not simple. As a consequence of the complex sampling design, there is dependence among sampled units. Hence, classical Efron bootstrap does not work in the case under examination. Resampling schemes based on pseudo-populations will be developed, and their main justifications and properties will be shown. The approach used is of asymptotic nature, and parallels results obtained by Bickel and Friedman for the i.i.d. case. Main applications of theoretical results are devoted to the construction of confidence intervals for finite population parameters. Finally, computational issues will be discussed. In allegato la locandina con l'abstract e i riferimenti per partecipare al seminario in presenza e a distanza.
Addressing dataset shift in supervised classification via data perturbation
November 25, 2022, 12:00
n supervised classification, dataset shift occurs when for the units in the test set a change in the distribution of a single feature, a combination of features, or the class boundaries, is observed with respect to the training set. As a result, in real data applications, the common assumption that the training and testing data follow the same distribution is often violated. Dataset shift might be due to several reasons; the focus is on what is called “covariate shift”, namely the conditional probability p(y|x) remains unchanged, but the input distribution p(x) differs from training to test set. Random perturbation of variables or units when building the classifier can help in addressing this issue. Evidence of the performance of the proposed approach is obtained on simulated and real data.
Stein’s Method Meets Statistics: A Review of Some Recent Developments
November 18, 2022, 12:00
Stein’s method compares probability distributions through the study of a class of linear operators called Stein operators. While initially studied in the field of probability, Stein’s method has led to significant advances in theoretical statistics, computational statistics and machine learning in recent years. In this talk, I will present some of these recent developments and, in doing so, try to stimulate further research into the successful field of Stein’s method and statistics. The topics I shall discuss include (if the time permits) new insights into the finite-sample approximation of estimators (like maximum likelihood estimators), a measure of the impact of the prior choice in Bayesian statistics, tools to benchmark and compare sampling methods such as approximate Markov chain Monte Carlo, deterministic alternatives to sampling methods, parameter estimation and goodness-of-fit testing. This talk is based on a large collaborative effort with many co-authors.
On mixtures of linear quantile regressions for longitudinal and clustered data
November 11, 2022, 12:00
Quantile regression represents a well established technique for modelling data when the interest is on the effect of predictors on the conditional response quantiles. When responses are repeatedly collected over time, or when they are hierarchically nested, dependence needs to be properly considered. A standard way of proceeding is based on including higher level unit-specific random coefficients in the model. The distribution of such coefficients may be either specified parametrically or left unspecified. In the last case, it can be estimated non parametrically by using a discrete distribution defined on G locations. This may approximate the distribution of time-constant and/or time-varying random coefficients, leading to a static, dynamic, or mixed-type mixture of linear quantile regression equations. An EM algorithm and a block-bootstrap procedure are employed to derive parameter estimates and corresponding standard errors. Standard penalized likelihood criteria are used to identify the optimal number of mixture components. This class of models is described by using a benchmark dataset and employing the functions in the newly develop lqmix R package.
Causal effects of chemotherapy regimen intensity on survival outcome through Marginal Structural Models
Novembre 4, 2022, 12:00
As patients develop toxic side-effects, cancer treatment is adapted over time by either delaying or reducing the dosage of the next chemotherapy course. In this talk Marginal Structural Models in combination with InverseProbability-of-Treatment Weighted estimators to assess the causal effects of chemotherapy regimen modifications on survival outcome will be discussed. The focus is on the use of actual treatment data and Received Dose Intensity in contrast with the use of intended treatment regimen. The latter approach, known as Intention to treat, is very common but also very far from the everyday clinical practice. In this talk, I will discuss the confounding nature of toxic side-effects data and shows the damaging effect of not including toxicity in the analysis. The method developed is applied to the osteosarcoma randomised clinical trials BO03 and BO06 (EORTC 80861 and 80931).
Density modelling with Functional Data Analysis
October 28, 2022, 12:00
Recent technological advances have eased the collection of big amounts of data in many research fields. In this scenario, a useful statistical technique is density estimation which represents an important source of information. One dimensional density functions represent a special case of functional data subject to the constraints to be non-negative and with a constant integral equal to one. Because of these constraints, densities functions do not form a vector space and a naive application of functional data analysis (FDA) methods may lead to nonvalid estimates. To address this issue, two main strategies can be found in the literature. In the first, the probability density functions (pdfs) are mapped into a linear functional space through a suitably chosen transformation. Established methods for Hilbert space valued data can be applied to the transformed functions and the results are moved back into the density space by means of the inverse transformation. In the second strategy, probability density functions are treated as an infinite dimensional compositional data since they are part of some whole which only carry relative information. In this work, by means of a suitable transformation, densities are embedded in the Hilbert space of square integrable functions where standard FDA methodologies can be applied.
The three-sigma rule to define antibody positivity: is it a beauty or a beast?
October 14, 2022, 12:00
Many epidemiological studies aim to estimate the proportion of individuals currently or previously infected by a given microorganism. Given that an infection inevitably leads to an immune response, this estimation exercise often requires identifying individuals who reach a minimal level of microbe-specific antibodies in their serum. This threshold invariantly is defined by the three-sigma rule: mean plus three times the standard deviation from the hypothetical antibody-negative population. Notwithstanding not being linked to a specific parametric distribution, it has the most intuitive interpretation in the context of a normal distribution. I will then discuss the problems of estimation bias and apparent control of specificity arising from applying this rule to nonnormal distributions for the seronegative population. I will use public data on antibody testing against the SARS-CoV2 to illustrate these problems. We should finally ask ourselves whether the three-sigma rule is a beautiful statistical concept or, instead, a little beast hidden in antibody data analysis.
A general framework for implementing distances for categorical variables
June 17, 2022, 13:30
In many statistical methods, distance plays an important role. For instance, data visualization, classification and clustering methods require quantification of distances among objects. How to define such distance depends on the nature of the data and/or problem at hand. For distance between numerical variables, in particular in multivariate contexts, there exist many definitions that depend on the actual observed differences between values. It is worth underlining that often it is necessary to rescale the variables before computing the distances. Many distance functions exist for numerical variables. For categorical data, defining a distance is even more complex as the nature of such data prohibits straightforward arithmetic operations. Specific measures therefore need to be introduced that can be used to describe or study structure and/or relationships in the categorical data. In this paper, we introduce a general framework that allows an efficient and transparent implementation for distance between categorical variables. We show that several existing distances (for example distance measures that incorporate association among variables) can be incorporated into the framework. Moreover, our framework quite naturally leads to the introduction of new distance formulations as well.
Model-assisted indirect small area estimation
May 27, 2022, 12:00
Generalised regression is the most common design-based model-assisted method for estimation of population means and totals in practical survey sampling. However, it is often unacceptable in the context of small area estimation, where one is interested in population means and totals for a large number of areas (or domains) and the sample sizes are either small or non-existent in many of them. In this seminar, we discuss an approach to extend generalised regression from direct estimation for the whole population to indirect estimation of all the small area populations. This requires to trade variance off with bias and enables a practical methodology for estimation at the different aggregation levels, which is coherent numerically (self-benchmarking) as well as conceptually in terms of the design-based model-assisted inference outlook. Estimation can be conducted by means of an *extended* weighting system that has as many sets of weights as the number of small areas: each set produces the estimate for a domain mean of one or more survey variables of interest and is, in this sense, multipurpose.
Extending the boundaries of a macroeconometric model for Italian economy to inequality
May 20, 2022, 12:00
According to the growing debate on the beyond-GDP approach, a strand of literature explores how the traditional system on national account (SNA), that is the pillar for the GDP measurement, could be extended to account for some of the main themes related to well-being and sustainability. In this presentation we extend the macroeconometric model for Italy developed by Istat (MeMo-It) introducing an inequality measure in the consumption function. Empirical analysis shows that a positive income shock that increases aggregate consumption in the current year might be completely off-set by the negative effect of the increase in inequality that becomes effective in the next year. In this framework, the impact of the Italian “reddito di cittadinanza”, a policy measure aiming at reducing poverty, has been evaluated. According to the results obtained we support the idea that a step forward on wellbeing and sustainability could be realized starting from a structural macroeconometric approach.
Bayesian Statistics applied to Early “Oncology Drug Development”
May 13, 2022, 17:00
Oncology dose finding studies, in general, aim at determining the maximum tolerated dose (MTD) reflecting the desire to treat patients who have limited options under the assumption that higher drug doses will have better therapeutic activity. We are describing different methods (ie 3+3, mTPI, mTPI-2, and BLRM). This seminar will feature speakers from Pfizer Inc., to share their insights and the recent statistical innovations to address the challenges. In addition to safety evaluation, Early Sign of Efficacy (ESOE) is a critical step in all early clinical programs to extend the development of a molecule or not. Robust and consistent calculation of the probability of making the right decision is critical. Innovative methodologies are needed to optimize these calculations and ensure all molecules are assessed in the same way across the oncology portfolio. Case studies will be discussed for dose finding and the utilization of Bayesian statistics in ESOE evaluation.
Bayesian Inference In High-dimensional Spatial Statistics: Conquering New Challenges
May 6, 2022, 17:00
Geographic Information Systems (GIS) and related technologies such as remote sensors, satellite imaging and portable devices that are capable of collecting precise positioning information, even on portable hand-held devices, have spawned massive amounts of spatial-temporal databases. Spatial "data science" broadly refers to the use of technology, statistical methods, computational algorithms to extract knowledge and insights from spatially referenced data. Applications of spatial-temporal data science are pervasive in the natural and environmental sciences; economics; climate science; ecology; forestry; and public health. With the abundance of spatial BIG DATA problems in the sciences and engineering, GIS and spatial data science will likely occupy a central place in the data revolution engulfing us. This talk will discuss construction and implementation of scalable Gaussian processes and the importance of conjugate Bayesian models in carrying out Bayesian inference for spatially and temporally oriented massive data sets exhibiting complex dependencies in diverse applications. We will elucidate recent developments in Bayesian statistical science such as geosketching and predictive stacking that can harness high performance scientific computing methods for spatial-temporal BIG DATA analysis and emphasize how such methods can be implemented on modest computing architectures. The talk will include specific examples of Bayesian hierarchical modeling in Light Detection and Ranging (LiDAR) systems and other remote-sensed technologies; environmental sciences; and public health.
Spatial and functional data over non-Euclidean domains
April 29, 2022, 12:00
Recent years have seen an explosive growth in the recording of increasingly complex and high-dimensional data. Classical statistical methods are often unfit to handle such data, whose analysis calls for the definition of new methods merging ideas and approaches from statistics and applied mathematics. My talk will in particular focus on spatial and functional data defined over non-Euclidean domains, such as linear networks, two-dimensional manifolds and non-convex volumes. I will present an innovative class of methods, based on regularizing terms involving Partial Differential Equations (PDEs), defined over the complex domains being considered. These physics-informed regression methods enable the inclusion in the statistical model of the available problem specific information, suitably encoded in the regularizing PDE. The proposed methods make use of advanced numerical techniques, such as finite element analysis and isogeometric analysis. A challenging application to neuroimaging data will be illustrated.
Factor models with downside risk
Aprile 22, 2022, 12:00
We propose a conditional model of asset returns in the presence of common factors and downside risk. Specifically, we generalize existing latent factor models in three ways: we show how to estimate the threshold which identifies the 'disappointment' event triggering the bad state of the world; we permit different factor structures for asset returns in good and bad states; we show how to recover the observable factors' risk premia from the estimated latent ones in different states. The usefulness of the model is illustrated through two applications to cross-sections of asset returns in equity markets and other major asset classes. Paper link https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3937321
Marine litter in the North-Western Ionian sea – Data features and space-time modeling
April 8, 2022, 12:00
Marine litter has recently become a recognized global ecological concern, and its distribution and impacts on deep-sea habitats are under continuous investigation. Here we focus on marine litter data collected as a by-product of trawl fishery surveys regularly conducted at a local scale in the Mediterranean. Litter data are multivariate, have space-time structure, and are semi-continuous, i.e. they combine information on occurrence and conditional-to-presence abundance. Data on potential environmental drivers obtained by remote sensing or GIS technologies are also available with different spatial support. The modeling strategy is based on a two-part model that enables handling the zero-inflation problem and the spatial correlation characterizing the data. In the spirit of multi-species distribution models, we propose to jointly infer different litter categories in a Hurdle-model framework. The effects of potential environmental drivers and shared spatial effects linking abundances and probabilities of occurrences of litter categories are implemented via the SPDE approach in the computationally efficient INLA context. Results support the possibility of better understanding the spatio-temporal dynamics of marine litter in the study area.
How much evidence do you need? Data Science and Bayesian Statistics to inform Environmental Policy during the COVID-19 Pandemic
April 4, 2022, 14:00
In this talk, I will provide an overview of data science methods, including methods for Bayesian analysis, causal inference, and machine learning, to inform environmental policy. This is based on my work analyzing a data platform of unprecedented size and representativeness. The platform includes more than 500 million observations on the health experience of over 95% of the US population older than 65 years old linked to air pollution exposure and several confounders. Finally, I provide an overview of studies on air pollution exposure, environmental racism, wildfires, and how they also can exacerbate the vulnerability to COVID-19. Press Coverage • https://www.nytimes.com/2021/08/13/climate/wildfires-smoke-covid.html • https://www.nytimes.com/2020/04/07/climate/air-pollution-coronavirus-covid.html • https://www.nytimes.com/2020/12/07/climate/trump-epa-soot-covid.html?smid=tw-share • https://science.sciencemag.org/content/360/6388/473 • https://www.npr.org/sections/health-shots/2017/06/28/534594373/u-s-air-pollution-stillkills-thousands-every-year-study-concludes • https://www.statnews.com/2016/11/14/climate-change-agreements/ • https://news.harvard.edu/gazette/story/2016/08/smoke-waves-will-affect-millions-incoming-decades/ • https://sites.sph.harvard.edu/francesca-dominici/senator-cory-booker-talking-about-nejmstudy/
Measures of Interrater Agreement
March 25, 2022, 12:00
Agreement among ratings or measurements provided by several raters (humans or devices) is considered in education, biomedical sciences, and other disciplines. For instance, the agreement among ratings of educators who assess on a new rating scale the language proficiency of a corpus of argumentative texts is considered to test reliability of the scale, or the agreement among clinical diagnoses provided by physicians is analysed for identifying the best treatment for the patient. In all these applications, the main interest is to analyse interrater absolute agreement, that is the extent that raters assign the same (or very similar) values on the rating scale. Many indices of interrater agreement on a whole group of subjects (objects) have been proposed. Less frequently agreement on single subjects has been considered, in spite of the fact that this is useful, for example, to request the raters for a specific comparison on single cases in which agreement is poor. In the seminar, after a critical review of the most used indices of interrater agreement, new subject-specific and global measures of absolute agreement for ratings on different levels of scale are presented. Some applications will show the advantages of the indices proposed.
Unsupervised whole graph embedding methods and applications
March 18, 2022, 12:00
Networks represent a powerful model for problems in different scientific and technological fields, such as neuroscience, molecular biology, biomedicine, sociology, social network analysis, and political science. As the number of network applications increases, so does a need for novel data analysis techniques. In many applications, the analysis focuses on a single network to cluster or classify its nodes or predict pairs of nodes that will form a link. In this talk, we focus on problems where a network is a statistical unit, and the analysis regards whole networks rather than their parts. Methods for learning features on networks focus mainly on the neighborhood of nodes and edges. We review some of the existing methodologies and introduce Netpro2vec, an embedding framework based on representations of graphs based on empirical probability distributions. The goal is to use basic node descriptions other than the degree, such as those induced by the Transition Matrix and Node Distance Distribution, to describe the local and global characteristics of the networks. The framework is evaluated on synthetic and real biomedical network datasets and compared to well-known competitors. Finally, open problems and future research directions are highlighted.
Multimodal regression with circular data
March 4, 2022, 12:00
There is a diverse range of practical situations where one may encounter random variables which are not defined on Euclidean spaces, as it is the case for circular data. Circular measurements may be accompanied by other observations, either defined on the unit circumference or on the real line, and in such cases it may be of interest to model the relationship between the variables from a regression perspective. It is not infrequent that parametric models fail to capture the underlying model given their lack of flexibility, but it may also happen that the usual paradigm of (classical) mean regression. We will present in this talk some recent advances in nonparametric multimodal regression, showing an adaptation of the mean-shift algorithm for regression scenarios involving circular response and/or covariate. Real data illustrations will be also presented. This is a joint work with María Alonso-Pena.

2021


Challenges in emulating target trials
December, 14, 2021
The framework of target trial emulation (TTE) is increasingly adopted when researchers wish to address causal questions using observational data. TTE has multiple advantages, starting from the clarity of explicitly specifying the hypothetical target experimental trial for the questions of interest. However, because the data often arise from linked administrative databases that are not created for research purposes, their handling demands extreme care if biased conclusions are to be avoided. Two main sources of bias have been broadly recognised in the epidemiological literature: immortal time bias and inappropriate selection of comparative groups. This talk will focus on other challenges to emulating target trials which are not commonly aired, using two examples. Hernán et al. Specifying a target trial prevents immortal time bias and other self-inflicted injuries in observational analyses. Journal of Clinical Epidemiology, 2016: 79 (2016) 70e75 Hernán and Robins. Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available. American Journal of Epidemiology, 2016: 183, 758–764 Suissa. Immortal time bias in observational studies of drug effects. Pharmacoepidemiol Drug Safety, 2007: 241–9
Testing for the Rank of a Covariance Kernel
December 10, 2021
How can we discern whether the covariance of a stochastic process is of reduced rank, and if so, what its precise rank is? And how can we do so at a given level of confidence? This question is central to a great deal of methods for functional data, which require low-dimensional representations whether by functional PCA or other methods. The difficulty is that the determination is to be made on the basis of i.i.d. replications of the process observed discretely and with measurement error contamination. This adds a ridge to the empirical covariance, obfuscating the underlying dimension. We describe a matrix-completion inspired test statistic that circumvents this issue by measuring the best possible least square fit of the empirical covariance's off-diagonal elements, optimised over covariances of given finite rank. For a fixed grid of sufficiently large size, we determine the statistic's asymptotic null distribution as the number of replications grows. We then use it to construct a bootstrap implementation of a stepwise testing procedure controlling the family-wise error rate corresponding to the collection of hypotheses formalising the question at hand. Under minimal regularity assumptions we prove that the procedure is consistent and that its bootstrap implementation is valid. The procedure circumvents smoothing and associated smoothing parameters, is indifferent to measurement error heteroskedasticity, and does not assume a low-noise regime. Based on joint work with Anirvan Chakraborty.
Robust Statistics for (big) data analytics
December 3, 2021,
Data rarely follow the simple models of mathematical statistics. Often, there will be distinct subsets of observations so that more than one model may be appropriate. Further, parameters may gradually change over time. In addition, there are often dispersed or grouped outliers which, in the context of international trade data, may correspond to fraudulent behavior. All these issues are present in the datasets that are analyzed on a daily basis by the Joint Research Centre of the European Commission and can only tackled by using methods which are robust to deviations to model assumptions (see for example [6]). This distance between mathematical theory and data reality has led, over the last sixty years, to the development of a large body of work on robust statistics. In the seventies of last century, it was expected that in the near future any author of an applied article who did not use the robust alterative would be asked by the referee for an explanation [9]. Now, a further forty years on, there does not seem to have been the foreseen breakthrough into the wider scientific universe. In this talk, we initially sketch what we see as some of the reasons for this failure, suggest a system of interrogating robust analyses, which we call monitoring [5] and describe a series of robust and efficient methods to detect model deviations, groups of homogeneous observations [10], multiple outliers and/or sudden level shifts in time series ([8]). Particular attention will be given to robust and efficient methods (known as forward search) which enables to use a flexible level of trimming and understand the effect that each unit (outlier or not) exerts on the model (see for example [1], [2] [7]). Finally, we discuss the extension of the above methods to transformations and to the big data context. The Box-Cox power transformation family for non-negative responses in linear models has a long and interesting history in both statistical practice and theory. The Yeo-Johnson transformation extends the family to observations that can be positive or negative. In this talk, we describe an extended Yeo-Johnson transformation that allows positive and negative responses to have different power transformations ([4] or [3]). As an illustration of the suggested procedure, we analyse data on the performance of investment funds, 99 out of 309 of which report a loss. The problem is to use regression to predict medium term performance from two short term DSS - Dipartimento di Scienze Statistiche - www.dss.uniroma1.it indicators. It is clear from scatterplots of the data that the negative responses have a lower variance than the positive ones and a different relationship with the explanatory variables. Tests and graphical methods from our robust analysis allow the detection of outliers, the testing of the values of transformation parameters and the building of a simple regression model. All the methods described in the talk have been included in the FSDA Matlab toolbox freely donwloadable as a toolbox from Mathworks file exchange or from github at the web address https://uniprjrc.github.io/FSDA/ References [1] Atkinson, A. C. and Riani, M. (2000). Robust Diagnostic Regression Analysis. Springer– Verlag, New York. [2] Atkinson, A. C., Riani, M., and Cerioli, A. (2004). Exploring Multivariate Data with the Forward Search. Springer–Verlag, New York. [3] Atkinson, A. C., Riani, M., and Corbellini, A. (2020). The analysis of transformations for profit-and-loss data. Journal of the Royal Statistical Society: Series C (Applied Statistics), 69(2), 251–275. [4] Atkinson, A. C., Riani, M., and Corbellini, A. (2021). The BoxCox Transformation: Review and Extensions. Statistical Science, 36(2), 239 – 255. [5] Cerioli, A., Riani, M., Atkinson, A. C., and Corbellini, A. (2018). The power of monitoring: How to make the most of a contaminated multivariate sample (with discussion). Statistical Methods and Applications, 27, 559–666. https://doi.org/10.1007/s10260–017–0409–8. [6] Perrotta, D., Torti, F., Cerasa, A., and Riani, M. (2020). The robust estimation of monthly prices of goods traded by the European Union. Technical Report EUR 30188 EN, JRC120407, European Commission, Joint Research Centre, Publications Office of the European Union, Luxembourg. ISBN 978-92-76-18351-8, doi:10.2760/635844. [7] Riani, M., Atkinson, A. C., and Cerioli, A. (2009). Finding an unknown number of multivariate outliers. Journal of the Royal Statistical Society, Series B, 71, 447–466. [8] Rousseeuw, P., Perrotta, D., Riani, M., and Hubert, M. (2019). Robust monitoring of time series with application to fraud detection. Econometrics and Sttaistics, 9, 108–121. [9] Stigler, S. M. (2010). The changing history of robustness. The American Statistician, 64, 277–281. [10] Torti, F., Perrotta, D., Riani, M., and Cerioli, A. (2018). Assessing trimming methodologies for clustering linear regression data. Advances in Data Analysis and Classification, 13, 227–257.

2019


seminar
22 Febbraio 2019 - Sala 34 ore 11

Giorgio Consigli (Università degli Studi di Bergamo) - Asset-liability management for occupational pension funds under market and longevity risk: a case study and alternative modelling approaches
22 Marzo 2019 - Aula V ore 15
The modelling of institutional ALM problems has a long history in stochastic programming starting in the late 80’s with the first industry developments such as the well-known Yasuda Kasai model (Ziemba, Turner, Carino et al, 1994) specifically for pension fund management (PF ALM). Due to economic and demographic pressures in most OECD countries and an increasing interest on PF ALM developments by the industry and by policy makers, we witness now-a-day a growing demand for R&D projects to the scientific community. Taking the view of a PF manager, the presentation will develop around the definition of a generic pension fund (PF) asset-liability management (ALM) problem and analyse the key underlying methodological implications of: (i) it's evolution from an early stage multistage stochastic programming (MSP) with recourse to most recent MSP and distributionally robust (DRO) formulations, (ii) a peculiar and rich risk spectrum including market risk as well as liability risk, such as longevity risk and demographic factors leading to (iii) valuation or pricing approaches based on incomplete market assumptions and, due to recent International regulation, (iv) a risk-based capital allocation for long-term solvency. The above represent fundamental stochastic and mathematical problems of modern financial optimisation. Two possible approaches to DRO are considered, based on a stochastic control framework or by explicitly introducing an uncertainty set for probability measures and formulating the inner DRO problem as a probability distance minimization problem over a given space of measures. Keywords: asset-liability management, multistage stochastic programming, distributional uncertainty, distributionally robust optimization, solvency ratio, liability pricing, longevity risk, capital allocation.
Gianluca Mastrantonio (Politecnico di Torino) - New formulation of the logistic-normal process to analyze trajectory tracking data
28 Gennaio 2019 - Sala 34 ore 10.30
Improved communication systems, shrinking battery sizes and the price drop of tracking devices have led to an increasing availability of trajectory tracking data. These data are often analyzed to understand animals behavior using mixture-type model. In this work, we propose a new model based on the logistic-normal process. Due to a new formalization and the way we specify the core- gionalization matrix of the associated multivariate Gaussian process, we show that our model, differently from other proposals, is invariant with respect to the choice of the reference element and of the order- ing of the components of the probability vectors. We estimate the model under a Bayesian framework, using an approximation of the Gaussian process needed to avoid impractical computational time. We perform a simulation study with the aim of showing the ability of the model to retrieve the parameters used to simulate the data. The model is then applied to the real data where a wolf is observed before and after procreation. Results are easy to interpret, showing differences in the two phases. Joint work with: Enrico Bibbona (Politecnico di Torino), Clara Grazian (Università di Pescara), Sara Mancinelli (università "Sapienza" di Roma)
Overdispersed-Poisson Model in Claims Reserving: Closed Tool for One-Year Volatility in GLM Framework
29 Marzo 2019 - Aula V ore 14.15

seminar
16-17 Aprile 2019 - Sala 34 ore 10-14

seminar
5 Giugno 2019 - Aula Master (Viale Regina Elena 295) ore 10

seminar
5 Giugno 2019 - Aula Master (Viale Regina Elena 295) ore 10

Daniel K. Sewell (University of Iowa) - An introduction to the statistical analysis of network data
9 e 10 Settembre 2019 - Aula VII (ex Castellano) ore 10-16 (con pausa)

Multi-study factor analysis for biological data
14 Novembre 2019 - Aula XIV (palazzina Tumminelli) ore 12
We introduce a novel class of factor analysis methodologies for the joint analysis of multiple studies. The goal is to separately identify and estimate 1) common factors shared across multiple studies, and 2) study-specific factors. We develop a fast Expectation Conditional-Maximization algorithm for parameter estimates and we provide a procedure for choosing the common and specific factors. We present simulations evaluating the performance of the method and we illustrate it by applying it to gene expression data in ovarian cancer and to nutrient-based dietary patterns and the risk of head and neck cancer. In both cases, we clarify the benefits of a joint analysis compared to the standard factor analysis. Moreover, we generalize the model in a Bayesian framework. We implement it using sparse modeling of high-dimensional factor loadings matrices, both common and specific, using the infinite gamma shrinkage prior. We propose a computationally efficient algorithm, based on a traditional Gibbs sampler, to produce the Bayes estimates of the parameters and to select the number of relevant common factors. We assess the operating characteristics of our method by means of simulation studies, and we present an application to the prediction of the biological signal from four gene expression studies on ovarian cancer.
Garyfallos Konstantinoudis - Discrete versus continuous domain models for disease mapping and applications on childhood cancers
22 Novembre 2019 - Sala 34 ore 12
The main goals of disease mapping is to calculate disease risk and identify high-risk areas. Such analyses are hampered by the limited geographical resolution of the available data. Typically data are counts of cases per spatial unit and the most common approach is the Besag-York-Molli ́e model (BYM). Less frequently, exact geocodes are available, allowing modelling a disease as a point process through Log-Gaussian Cox processes (LGCPs). The objective of this study is to examine in a simulation the performance of BYM and LGCPs for disease mapping. We simulated data in the Canton of Zurich in Switzerland sampling cases from the true population mimicking the childhood leukaemia incidence (n=334 during 1985-2015). We considered 39 different scenarios varying in the risk generating function (step-wise, smooth, flat risk), the size of the high-risk areas (1, 5 and 10km radii), the risk increase within the high-risk areas (2 and 5-fold) and the number of cases (n, 5n and 10n). We used the root mean integrated square error (RMISE) to examine the ability of the models to recover the true risk surface and their sensitivity/specificity in identifying highrisk areas. We found that, for larger radii, LGCPs recover the true risk surface with lower error across almost all scenarios (median RMISE: 9.17-27.0) compared to the BYM (median RMISE: 9.12-35.6). For radii = 1km and flat risk surfaces BYM performs better. In terms of sensitivity and specificity across almost all scenarios the median area under the curve (AUC) for LGCPs was higher (median AUC: 0.81-1) compared to the BYM (median AUC: 0.65-0.93). We applied these methods to childhood leukaemia incidence in the canton of Zurich during 1985-2015 and identified two high-risk spatially coherent areas. Our findings suggest that there are important gains to be made from the use of LGCP models in spatial epidemiology.
Roberta Varriale (ISTAT) - Machine learning methods for estimating the employment status in Italy
29 Novembre 2019 - Sala 34 ore 14
In recent decades, National Statistical Institutes have focused on producing official statistics by exploiting multiple sources of information (multi-source statistics) rather than a single source, usually a statistical survey. The growing importance of producing multi-source statistics in official statistics has led to increasing investments in research activities in this sector. In this context, one of the research projects addressed by the Italian National Statistical Institute (Istat) concerned the study of methodologies for producing estimates on employment rates in Italy through the use of multiple sources of information, survey data and administrative sources. The data comes from the Labour Force (LF) survey conducted by Istat and from several administrative sources that Istat regularly acquires from external bodies. The “quantity” of information is very different: those coming from administrative sources concern about 25 million individuals, while those coming from the LF survey refer to an extremely limited number (about 330,000) of individuals. The two measures do not agree on employment status for about 6% of the units from the LF survey. One proposed approach uses a Hidden Markov model to take into account the deficiencies in the measurement process of both survey and administrative sources. The model describes a measurement process as a function of a time-varying latent state (in this case the employment category), whose dynamics is described by a Markov chain defined over a discrete set of states. At present, the implementation phase for the production process of statistics on employment through the use of HM models is coming to an end in Istat. The present work describes the use of Machine Learning methods to predict the individual employment status. This approach is based on the application of decision tree and random forest models, that are predictive models, usually used to classify instances of large amounts of data. In the work, the obtained results will be described, together with their usefulness in this application context. The models have been applied through the use of the software R.

2018


Yves Tillé (Université de Neuchatel) - How to select a sample?
27 Novembre 2018 - Sala 34 ore 14.30
The principles of sampling can be synthesized in randomization, restriction and over-representation. Define a sample design – define stratification, equal/unequal selection probability, etc. – means to use prior information and it is equivalent to assume a model on the population. Several well-known sampling designs are optimal related to models that maximizes the entropy. In the Cube method the prior information are used to derive a sample that match the total or means of auxiliary variables. In this respect, the sample is called balanced. Furthermore, if distances between statistical units – based on geographical coordinates or defined via auxiliary variables – are available, it could be interesting to spread the sample in the space in order to make the design more efficient. In this perspective, new spatial sampling methods, such as the GRTS, the local pivotal method and the local cube, will be covered.

© Università degli Studi di Roma "La Sapienza" - Piazzale Aldo Moro 5, 00185 Roma