Previous statistics seminars 2023
Autumn
Seminar 2023-11-22: Spatial extreme values based on the multivariate exponential power distribution
Speaker: Alexander Engberg, Department och Statistics, Uppsala University
Time and place: 2023-11-22 at 10:15-11:30, Ekonomikum, Room H317
Abstract
We show how the multivariate exponential power distribution can provide a flexible model option for spatial extreme values from asymptotically independent processes. The distribution is a generalization of the multivariate Gaussian distribution and can be expressed as a Gaussian scale mixture. Furthermore, it allows for heavy tails, where the heaviness of the joint tail is estimated from the data. We derive the joint tail decay rate of the distribution, and show how to perform conditional simulations from it. The distributions practical use is illustrated by comparing it to the Gaussian model when modelling wind gust data gathered in the central eastern parts of Sweden. (Joint work with Lukas Arnroth)
Seminar 2023-10-18: Sensitivity analysis in multilevel modelling when data is missing not at random
Speaker: Minna Genbäck, Department of Statistics, Umeå University
Time and place: 2023-10-18 at 10:15-11:30, Ekonomikum room, F416
Abstract
In this study we develop a sensitivity analysis method for missing not at random data when using multilevel models (mixed effects models). By allowing for milder assumptions than missing at random we do not fully identify the parameters of interest, however we are able to partially identify them and thereby derive bounds for estimates and corresponding inference. We apply the method in an analysis of the association between loneliness and physical activities with memory trajectories while adjusting for different demographic, socioeconomic and health variables. The talk will include an introduction to this type of sensitivity analysis with simpler models (e.g. linear and probit regression).
Seminar 2023-10-04: Model Instabilities for Structured Multivariate Data
Speaker Matúš Maciak, Department of Probability & Mathematical Statistics, Charles University, Prague, Czech Republich
Time and place: 2023-10-04 at 10:15-11:30, Ekonomikum, Room H317
Spring 2023
2023-05-24 (review seminar): Posterior rate of convergence for composite quantile regression
Speaker: Lukas Arnroth, Department of Statistics, Uppsala University. Time and place: 2023-05-24 at 10:15–11:30, Ekonomikum room H317.
Abstract
Composite quantile regression is based on the convex combination of single quantile quantile loss functions and enjoys many advantages over single quantile regression. The Bayesian extension is based on the finite mixture of asymmetric Laplace densities. This paper mainly aims to contribute to the theoretical justification of Bayesian composite quantile regression from the perspective of Bayesian density estimation. As such, we further show that the asymmetric Laplace distribution can be used for Bayesian density estimation in general. We obtain upper bounds on rates of convergence for mixtures of asymmetric Laplace densities. For finite mixtures we obtain the parametric rate up to a logarithmic factor, and a slower rate for infinite mixture.
2023-05-17 (review seminar): Can Model Averaging Improve Propensity Score Based Estimation of Average Treatment Effects?
Speaker: Valentin Zulj, Ddepartment of Statistics, Uppsala University. Time and place: 2023-05-17 at 10:15–11:30, Ekonomikum room H317.
Abstract
In drawing causal inferences from observational data, researchers often need to model the propensity score. To date, the literature on the estimation of propensity scores is vast, and includes covariate selection algorithms as well as super learners and model averaging procedures. The latter often tune the estimated scores to be either very accurate or to provide the best possible result in terms of covariate balance.
This paper focuses on using inverse probability weighting to estimate average treatment effects, and makes the assertion that the context requires both accuracy and balance to yield suitable propensity scores. Using Monte Carlo simulation, the paper studies whether frequentist model averaging can be used to simultaneously account for both balance and accuracy in order to remove some bias from estimated treatment effects. The simulations suggest that the combined procedure does not result in a consistent or substantial reduction of bias.
2023-05-03: Penalized QMLE with parameters on the boundary and an application to the class of ARCH(Q) for large Q
Speaker: Anders Rahbek, University of Copenhagen, Denmark. Time and place: 2023-05-03 at 10:15–11:30, Ekonomikum room H317.
2023-04-19: Relative survival and other summary measures of survival useful for population-based cancer data
Speaker: Therese Andersson, Karolinska institutet (KI), Stockholm. Time and place: 2023-04-19 at 10:15–11:30, Ekonomikum room H317.
Abstract
I will introduce the field of population-based cancer survival analysis and its role in cancer control. I will especially cover the concept of relative survival and why it is often preferred over cause-specific survival for the study of cancer patient survival using data collected by population-based cancer registers. I will also present different summary measures of cancer-patient survival, such as, the loss in life expectancy due to cancer and crude vs net probabilities of death. Each of these measures show different aspects of cancer patient survival, and examples from published population-based studies will be presented and discussed.
2023-04-12: Cross-Lingual Dependency Parsing
Speaker: Sara Stymme, Department of Linguistics and Philology, Uppsala University. Time and place: 2023-04-12 at 10:15–11:30, Ekonomikum room H317.
Abstract
Lately, there has been an increasing amount of work on cross-lingual learning, and how models for a target language can be improved using data from other languages. In this talk, I will focus on dependency parsing, the task of constructing a syntactic tree for an input sentence.
The resource Universal Dependencies, with harmonised treebanks for more than 100 languages, serves as a great test bed for cross-lingual learning. I will show how we can improve results for a specific language, by including training data from one or more other languages. This is useful particularly for low-resource languages, but I will show that it can be useful also for high-resource languages, especially when there is no in-language data for the target domain. I will also discuss how to choose suitable transfer languages for a given target language.
2023-03-22: Variational inference of dynamic factor models with arbitrary missing data
Speaker: Erik Spånberg, Department of Statistics, Stockholm University. Time and place: 2023-03-22 at 10:15–11:30, Ekonomikum room H317.
Abstract
Many forecasting institutions deal large multifaceted data sets, with time series of different frequencies, sample sizes, publication dates and general availability patterns. The same institutions are often under crucial time constraints, limiting their analytical capabilities. Dynamic factor models (DFMs) are popular tools for analysing large data sets, however they are often estimated by point-estimate methods, disregarding parameter uncertainty. Parameter uncertainty can be addressed by Bayesian inference, but that may be too computationally costly. Variational inference is a method that approximates Bayesian inference. I show that it can be applied to DFMs – including arbitrary availability patterns in the data – with large computational gains. This allows for deeper and more versatile analysis for the most fast-paced forecasting institutions.
2023-03-08: Distances for gene-expression transcriptomics: Metric clustering of mRNA transcripts
Speaker: Jim Blevins, Department of Statistics, Uppsala University. Time and place: 2023-03-08 at 10:15–11:30, Ekonomikum room H317.
Abstract
In molecular-biology, the expression of genes is measured for hundreds or thousands of genes for each observational unit (tissue or microbes). Genes with similar vectors of gene expression (mRNA transcripts or transcriptomics) are paired; gene-pair similarity is quantified using metrics on mRNA vectors. Using such distances on paired genes, biostatisticians identify gene sets (and gene modules) for further investigation. We develop metrics for the identification of interesting gene sets. We relate our metric approach to statistical methods --- e.g., the least-squares methods of Cramér, Wold, and Whittle.
2023-02-08: Modelling brain connectivity using hierarchical VAR models
Speaker: Anders Lundquist, Umeå University. Time and place: 2023-02-08 at 10:15–11:30, Ekonomikum room H317.
Abstract
Analysis of brain connectivity is important for understanding how information is processed by the brain. We propose a hierarchical vector autoregression (VAR) model for analysing brain connectivity, modelling so-called functional and effective connectivity simultaneously and allows for both group- and single-subject inference as well as group comparisons. We illustrate our approach in a resting-state fMRI data set with autism spectrum disorder (ASD) patients and healthy controls, and compare with similar models used in existing connectivity literature.
In the talk, I will spend a fair amount of time introducing the application before diving into the statistical issues. This is joint work with Bertil Wegmann (LiU), Anders Eklund (LiU), and Mattias Villani (LiU/SU).