Palestrantes

As palestras serão proferidas em português, espanhol ou inglês.

Palestra de abertura (quinta às 9:30) p-value and e-value: Are they diferent in practice? Carlos Alberto de Bragança Pereira (IME–USP). Abstract: e-values are used in Bayesian significance tests for sharp hypotheses. As the p-values, they intend to measure the consistency of data with the null hypothesis. For multinomial categorical datum from 2 by 2 contingency tables, functions relating these two measures are presented for standard situations: McNemar, Homogeneity, Independence and Hardy-Weinberg Equilibrium.

Palestra 1 (quinta às 10:25) Coherency and sets of full conditional measures. Fabio Cozman (Poli–USP). Abstract: Requirements of coherency imposed by de Finetti lead to his "Fundamental Theorem": assessment of some previsions imply bounds on other previsions. Similar coherency arguments have been advanced by many in connection with conditional previsions. We ask, What happens if assessments are interval-/set-valued, due to imprecision, incompleteness, disagreement, or simply due to lack of resources for elicitation? This talk reviews recent results on theories of coherent behavior that extend de Finetti's original ideas so as to handle interval-/set-valued versions of full conditional measures.

Palestra 2 (quinta às 11:20) Decoupling, Randomization and Objective Bayesian Inference. Julio Michael Stern (IME–USP). Abstract: We analyze the role of randomization in (Bayesian) Statistics. We review the decoupling principle, with examples in Physics and other sciences. We also review several standard decoupling operators of linear algebra, like LU, Cholesky and SVD factorizations. Cholesky and sparse Cholesky factorizations are presented as the decoupling operators of choice for covariance models. We also show that, from an abstract combinatorial perspective, almost identical decoupling operators arise in the theory of Bayesian networks. Finally, we discuss the role of randomization in the induction of decoupled models, and review the work of C.S.Peirce from an historical perspective.

Palestra 3 (quinta às 14:00) Multigrid Priors and fMRI. Nestor Caticha (IF–USP). Abstract: We present a non parametric Bayesian multi scale method to characterize the Hemodynamic Response HR as function of time. This is done by extending and adapting the Multigrid Priors (MGP) method proposed in (Amaral, Caticha). We choose an initial HR model and apply the MGP method to assign a posterior probability of activity for every pixel. This can be used to construct the map of activity. But it can also be used to construct the posterior averaged time series activity for different regions. This permits defining a new model which is only data dependent. Now in turn it can be used as the model behind a new application of the MGP method to obtain another posterior probability of activity. The method converges in just a few iterations and is quite independent of the original HR model, as long as it contains some information of the activity/rest state of the patient. We apply this method of HR inference both to simulated and real data of blocks and event-related experiments. Receiver operating characteristic (ROC) curves are used to measure the number of errors with respect to a few hyperparameters. We also study the deterioration of the results for real data, under information loss. This is done by decreasing the signal to noise ratio and also by decreasing the number of images available for analysis and compare the robustness to other methods.

Palestra 4 (quinta às 15:00) Adaptive Intervals for Hypothesis Testing (aiht): A Synthesis of Procedures. Luis Raúl Pericchi (University of Puerto Rico). Abstract: Very little has been proposed in the literature, which may be teachable in elementary courses, to bridge above the main disagreement between Schools of Statistics, and in particular to alleviate the discrepancy between Significance Hypothesis Testing (and Testing via Intervals with Fixed Size) and Bayesian Hypothesis Testing. It is well known, that huge samples lead to an almost sure rejection of Null Hypothesis by Classical Hypothesis Testing and by Bayesian Intervals with fixed probability that do not adapt with the amount of information and sample size. We propose here a calibration of p-values and significance levels that convey specific guidelines in how to diminish the alpha-levels (or increase the posterior probability of a Bayesian Interval) as the sample sizes grows. This effectively alleviates the discrepancy of Bayes Factors and Classical Testing, and makes probability intervals apt for performing a hypothesis test with large sample sizes. Furthermore, the resulting adaptive alpha levels imply enlarged intervals that can be used for testing hypothesis with similar decisions for both Schools of Statistics. Our basic position is that the posterior probability of a model has the same interpretation for any sample size, but that a p-value, or an alpha-level is heavily dependent on the sample size and should be corrected in order to be close to a posterior model probability. Once this is done, the "disagreement" in Statistics, largely disappears.

Palestra 5 (quinta às 16:30) Reliability nonparametric Bayesian estimation in parallel systems. Adriano Polpo (UFSCar). Abstract: Nonparametric Bayesian estimators of all survive distribution functions involved in a parallel system are presented. The inference problem starts when one observes the failure time of the machine and identifies the last component failing at the time of observation. Dirichlet multivariate processes forming a class of prior distributions are considered for nonparametric Bayes estimation for components distribution functions and the reliability of the system. The only restriction on the component sets of discontinuity points is that they are disjoint sets. For illustration two numerical examples are presented.

Palestra 6 (sexta às 9:00) Densities of Regular Variation for Saint Petersburg Envelopes. Sergio Wechsler (IME–USP). Abstract: The Two Envelopes Paradox is solved by correct conditioning whenever the expected amount of money placed in the envelopes is finite. We argue however that the paradox can still remain if the expected amount is not finite. Densities yielding such – a Saint Petersburg – situation are seen to be essentially of regular variation.

Mesa Redonda (sexta às 10:30) Papel da estatística na medicina e na sociedade. Medicina baseada em evidências, revisões sistemáticas e meta-análise. Hélio Elkis. (Instituto de Psiquiatria / HCFMUSP). O papel dos clinical trials no desenvolvimento de novos medicamentos. Sonia Dainesi. (Núcleo de Apoio à Pesquisa Clínica – Diretoria Clínica / HCFMUSP). O desenvolvimento da estatística Bayesiana em Medicina. Basílio B. Pereira. (Faculdade de Medicina – COPPE &ndash HUCFF / UFRJ)

Palestra 7 (sexta às 14:00) Extensions of Bayesian Networks. Cassio de Campos (EACH-USP). Abstract: This talk explores extensions of Bayesian Networks, such as credal networks and semi-qualitative probabilistic networks. The former extends usual probabilistic networks to sets of probability measures, while the latter combines numeric and qualitative information in the same framework. We show that inferences are hard, but can be dealt through multilinear programming. We then discuss learning and describe a Bayesian-minded method that employs the Imprecise Dirichlet Model to generate set-valued estimates.

Palestra 8 (sexta às 14:50) Testing and estimating the non-disjunction fraction in meiosis I using reference priors. Rosangela H. Loschi (UFMG). Abstract: In this paper we analyze the fraction of non-disjunction in Meiosis I assuming reference (non-informative) priors. We consider Jeffreys's approach to built a non-informative prior (Jeffreys's prior) for the fraction of non-disjunction in Meiosis I. We prove that Jeffreys's prior is a proper distribution. We perform Monte Carlo studies in order to compare Bayes estimates obtained assuming Jeffreys's and the uniform priors. We consider full Bayesian significance test (FBST) and Bayes factor (BF) for testing precise hypothesis on the fraction of non-disjunction in Meiosis I. The ultimate goal of this paper is to compare these two test procedures througth simulation studies using both prior specifications. An application to Down Syndrome data is also presented.

Palestra 9 (sexta às 16:00) How the updating process can be fully characterized? Identifiability in the Bayesian approach. Ernesto San Martin (PUC, Chile). Abstract: Some Bayesian literature states that the lack of identifiability is not a real problem for the Bayesian inference. In this talk, we want to evaluate such an statement. To this end, we first discuss the following question: how identifiability in a Bayesian set-up can be defined? Secondly, we will show how works such a concept in a very simple statistical model, namely a normal distribution the mean of which is a sum of two mutually prior independent random variables. The objective of this example is to show what is the impact of unidentifiability at the inference level. Third, we will study the identifiability of semiparametric IRT-type models. In this context, it will be discussed what is the eventual empirical meaning of a semiparametric model.

Palestra 10 (sexta às 16:50) Statistical analysis of medical data: the Bayesian revolution. Jorge Achcar (FMRP–USP). Abstract: In this talk, we will present some examples of Bayesian analysis for medical data. The use of recent stochastic simulation methods in Bayesian statistics has been a great revolution for statistical data analysis, especially considering medical data. Usually, medical data could have small sample sizes, missing data, censored data and complex modelling structures that could invalidate the usual existing classical inference methods to analyse the data. The use of recent Bayesian software, like the WinBugs software gives a great simplification for the consulting statistical work in a medical school. We will present three special applications analysed in CEMEQ (Centro de Métodos Quantitativos) of the medical school of Ribeirão Preto, University of São Paulo: a application with correlated binary data in the presence of covariates; a application with repeated measures with a Cushing disease data set and a application with medical diagnostic test evaluation in the presence of verification bias.

Palestra 11 (sexta às 17:40) The Tempo and Mode of Evolution of Transposable Elements as Revealed by Molecular Phylogenies Reconstructed from Mosquito Genomes. Claudio Struchiner (Fiocruz). Abstract: We here present estimates of key parameters guiding transposable elements (TE) invasion dynamics as revealed by molecular phylogenies reconstructed from Anopheles gambiae and Aedes aegypti mosquito genome projects. Our analysis follows four steps: (i) mining the two mosquito genomes currently available in search of TE families; (ii) fitting, to selected families found in (i), a phylogeny tree under the general time-reversible (GTR) nucleotide substitution model with an uncorrelated lognormal relaxed clock (UCLN) and a non-parametric demographic model; (iii) fitting a non-parametric coalescent model to the tree generated in (ii); (iv) fitting parametric models motivated by ecological theories to the curve generated in (iii). The demographic component implied by this approach is of great epidemiological interest since it can help in identifying the molecular and ecological approximate conditions under which a transposable elements will spread through a host population. By exploring the analytic potential of coalescence models fitted to data generated by genome projects, we address the lack of empirical data against which putative models developed in the past could be fitted and checked.