McGill Statistics Seminar

Date: 2016-03-18

Time: 15:30-16:30

Location: BURN 1205

Abstract:

We consider the problem of estimating the mean vector, q, of a multivariate spherically symmetric distribution under a loss function which is a concave function of squared error. In particular we find conditions on the shrinkage factor under which Stein-type shrinkage estimators dominate the usual minimax best equivariant estimator. In problems where the scale is known, minimax shrinkage factors which generally depend on both the loss and the sampling distribution are found. When the scale is estimated through the squared norm of a residual vector, for a large subclass of concave losses, we find minimax shrinkage factors which are independent of both the loss and the underlying distribution. Recent applications in predictive density estimation are examples where such losses arise naturally.

Date: 2016-03-11

Time: 15:30-16:30

Location: BURN 1205

Abstract:

We consider the problem of learning the structure of a non-Gaussian graphical model. We introduce two strategies for constructing tractable nonparametric graphical model families. One approach is through semiparametric extension of the Gaussian or exponential family graphical models that allows arbitrary graphs. Another approach is to restrict the family of allowed graphs to be acyclic, enabling the use of fully nonparametric density estimation in high dimensions. These two approaches can both be viewed as adding structural regularization to a general pairwise nonparametric Markov random field and reflect an interesting tradeoff of model flexibility with structural complexity. In terms of graph estimation, these methods achieve the optimal parametric rates of convergence. In terms of computation, these methods are as scalable as the best implemented parametric methods. Such a “free-lunch phenomenon” makes them extremely attractive for large-scale applications. We will also introduce several new research directions along this line of work, including latent-variable extension, model-based nonconvex optimization, graph uncertainty assessment, and nonparametric graph property testing.

Date: 2016-02-26

Time: 15:30-16:30

Location: BURN 1205

Abstract:

In this talk, we will consider a portfolio of dependent risks represented by a vector of dependent random variables whose joint cumulative distribution function (CDF) is defined with an Archimedean copula. Archimedean copulas are very popular and their extensions, nested Archimedean copulas, are well suited for vectors of random vectors in high dimension. I will describe a simple approach which makes it possible to compute the CDF of the sum or a variety of other functions of those random variables. In particular, I will derive the CDF and the TVaR of the sum of those risks using the Frank copula, the Shifted Negative Binomial copula, and the Ali-Mikhail-Haq (AMH) copula. The computation of the contribution of each risk under the TVaR-based allocation rule will also be illustrated. Finally, the links between the Clayton copula, the Shifted Negative Binomial copula, and the AMH copula will be discussed.

Date: 2016-02-19

Time: 15:30-16:30

Location: BURN 1205

Abstract:

The study of convergence of random walks to well defined curves is founded in the fields of complex analysis, probability theory, physics and combinatorics. The foundations of this subject were motivated by physicists interested in the properties of one-dimensional models that represented some form of physical phenomenon. By taking physical models and generalizing them into abstract mathematical terms, macroscopic properties about the model could be determined from the microscopic level. By using model specific objects known as observables, the convergence of the random walks on particular lattice structures can be proven to converge to continuous curves such as Brownian Motion or Stochastic Loewner Evolution as the size of the lattice step approaches 0. This seminar will introduce the field of statistical lattice models, the types of observables that can be used to prove convergence as well as a proof for the q-state Potts model showing that local non-commutative matrix observables do not exist. No prior physics knowledge is required for this seminar.

Date: 2016-02-05

Time: 15:30-16:30

Location: BURN 1214

Abstract:

Estimating causal exposure effects in observational studies ideally requires the analyst to have a vast knowledge of the domain of application. Investigators often bypass difficulties related to the identification and selection of confounders through the use of fully adjusted outcome regression models. However, since such models likely contain more covariates than required, the variance of the regression coefficient for exposure may be unnecessarily large. Instead of using a fully adjusted model, model selection can be attempted. Most classical statistical model selection approaches, such as Bayesian model averaging, do not readily address causal effect estimation. We present a new model averaged approach to causal inference, Bayesian causal effect estimation (BCEE), which is motivated by the graphical framework for causal inference. BCEE aims to unbiasedly estimate the causal effect of a continuous exposure on a continuous outcome while being more efficient than a fully adjusted approach.

Date: 2016-01-29

Time: 15:30-16:30

Location: BURN 1205

Abstract:

In this talk, we investigate the problem of estimating high-dimensional networks in which there are a few highly connected “hub" nodes. Methods based on L1-regularization have been widely used for performing sparse selection in the graphical modelling context. However, the L1 penalty penalizes each edge equally and independently of each other without taking into account any structural information. We introduce a new method for estimating undirected graphical models with hubs, called the hubs weighted graphical lasso (HWGL). This is a two-step procedure with a hub screening step, followed by network reconstruction in the second step using a weighted lasso approach that incorporates the inferred network topology. Empirically, we show that the HWGL outperforms competing methods and illustrate the methodology with an application to microbiome data.

Date: 2016-01-22

Time: 15:30-16:30

Location: BURN 1205

Abstract:

Influential units are those which make classical estimators (e.g., the Horvitz-Thompson estimator or calibration estimators) very unstable. The problem of influential units is particularly important in business surveys, which collect economic variables, whose distribution are highly skewed (heavy right tail). In this talk, we will attempt to answer the following questions:

(1) What is an influential value in surveys? (2) How measure the influence of unit? (3) How reduce the impact of influential units at the estimation stage?

Date: 2015-11-13

Time: 15:30-16:30

Location: BURN 1205

Abstract:

Logistic or other constraints often preclude the possibility of conducting incident cohort studies. A feasible alternative in such cases is to conduct a cross-sectional prevalent cohort study for which we recruit prevalent cases, i.e., subjects who have already experienced the initiating event, say the onset of a disease. When the interest lies in estimating the lifespan between the initiating event and a terminating event, say death for instance, such subjects may be followed prospectively until the terminating event or loss to follow-up, whichever happens first. It is well known that prevalent cases have, on average, longer lifespans. As such, they do not form a representative random sample from the target population; they comprise a biased sample. If the initiating events are generated from a stationary Poisson process, the so-called stationarity assumption, this bias is called length bias. I present the basics of nonparametric inference using length-biased right censored failure time data. I’ll then discuss some recent progress and current challenges. Our study is mainly motivated by challenges and questions raised in analyzing survival data collected on patients with dementia as part of a nationwide study in Canada, called the Canadian Study of Health and Aging (CSHA). I’ll use these data throughout the talk to discuss and motivate our methodology and its applications.

Date: 2015-11-06

Time: 15:30-16:30

Location: BURN 1205

Abstract:

Most regression models in biostatistics assume identifiability, which means that each point in the parameter space corresponds to a unique likelihood function for the observable data. Recently there has been interest in Bayesian inference for non-identifiable models, which can better represent uncertainty in some contexts. One example is in the field of epidemiology, where the investigator is concerned with bias due to unmeasured confounders (omitted variables). In this talk, I will illustrate Bayesian analysis of a non-identifiable model from epidemiology using government administrative data from British Columbia. I will show how to use the software STAN, which is new software developed by Andrew Gelman and others in the USA. STAN allows the careful study of posterior distributions in a vast collection of Bayesian models, including non-identifiable models for bias in epidemiology, which are poorly suited to conventional Gibbs sampling.

Date: 2015-10-23

Time: 15:30-16:30

Location: BURN 1205

Abstract:

Finite mixture regression models have been widely used for modeling mixed regression relationships arising from a clustered and thus heterogenous population. The classical normal mixture model, despite of its simplicity and wide applicability, may fail dramatically in the presence of severe outliers. We propose a robust mixture regression approach based on a sparse, case-specific, and scale-dependent mean-shift parameterization, for simultaneously conducting outlier detection and robust parameter estimation. A penalized likelihood approach is adopted to induce sparsity among the mean-shift parameters so that the outliers are distinguished from the good observations, and a thresholding-embedded Expectation-Maximization (EM) algorithm is developed to enable stable and efficient computation. The proposed penalized estimation approach is shown to have strong connections with other robust methods including the trimmed likelihood and the M-estimation methods. Comparing with several existing methods, the proposed methods show outstanding performance in numerical studies.