Search
Are you looking for internal information?
Showing 23781-23800 of 35043 hits
- Seminar: Jonathan Januar, Melbourne School of Psychological Sciences, University of Melbourne
Department of Statistics
Models for missing data in covert networks Seminar Date: Thursday 30 May 2024 Time: 13.00 – 14.00 Location: Campus Albano, Lecture room 30, house 4, level 2 Abstract Missing data are a prominent issue wherever data can be collected and is an especially notable problem when dealing with network data. As network data describe dependencies between nodes, missing data can lead to a multitude of issues ranging from inaccurate estimates to differing inferential claims. When considering the secretive aspect of covert networks - such as criminal or terrorist networks - validated and reliable data are very difficult to obtain and thus leads to great concerns with missing and unreliable data in the discipline. This project aims to discuss the considerations and nuances for addressing missing data in covert networks. Despite missing data being discussed in the social network analysis literature, there is not enough detail or discussion to explicitly describe the patterns by which missing data are missing. We explore various empirical covert network biases and represent them using explicit models for mechanisms by which missing data are generated. We further extend models for missingness to models for the sampling of edges in a graph in order to represent empirical biases in the process by which covert networks are observed. Simulations of these models are applied to empirical covert network datasets to evaluate the effects of specifying explicit patterns of missingness. Ultimately, this project aims to bridge disconnects between statistical assumptions involved in addressing missingness in network analysis and the harsh empirical data collection conditions of covert networks by introducing statistically principled methods of generating plausible missingness mechanisms.
- Seminar: Per Gösta Andersson, Department of Statistics, Stockholm University
Department of Statistics
Confidence interval construction for a binomial p Seminar Date: Wednesday 22 May 2024 Time: 13.00 – 14.00 Location: Campus Albano, Lecture room 14, house 2, level 2 Abstract The Wald confidence interval has for a long time been the standard interval for a binomial p. Regrettably, it is still very much in use even if, as Phil Kott, RTI, so neatly has expressed it: “Everyone knows that the Wald interval stinks”. We will take a look at what other colleagues have to say and also at how our course books are dealing with this problem. Besides a display of the bad properties of the Wald interval, better alternatives will be presented. References Andersson, P.G. (2022). Approximate confidence intervals for a binomial p --- once again. Statistical Science 37, 598--606. Andersson, P.G. (2023). The Wald confidence interval for a binomial p as an illuminating ``bad" example. The American Statistician 77, 443--448. Andersson P.G. (2024). Confidence Interval for a Binomial p: Recommend the Wilson interval to Students, Chance 37, 13—15. Brown, L.D., Cai, T.T. and DasGupta, A. (2001). Interval estimation for a binomial proportion. Statistical Science 16, 101--117. Brown, L.D., Cai, T.T. and DasGupta, A. (2002). Confidence intervals for a binomial proportion and asymptotic expansions. The Annals of Statistics 30, 160--201.
- Seminar: Gebrenegus Ghilagaber, Department of Statistics, Stockholm University
Department of Statistics
Bayesian Ordinal Modeling of Responses on the Likert-Scale: Illustration with Effects of Online Teaching on Students with Special Needs Seminar Date: Wednesday 22 November 2023 Time: 13.00 – 14.00 Location: Campus Albano, Lecture room 26, house 4, level 2 Abstract We demonstrate the power of Bayesian approach to modeling data with ordered outcome variable and illustrate it using data on the effect of the imposed online teaching during the Covid-19 pandemic on students with special needs. It is known that even frequentist ordinal modeling provides better theoretical interpretation and inference for such data than models which treat the ordered (Likert) outcome variable as metric. We argue that Bayesian approach adds flexibility in specifying the ordered models and provides goodness-of-fit measures to compare alternative models such as cumulative model and adjacent-category model with equal and unequal variances across levels of a covariate. Data for illustration come from about 6000 students who responded to a questionnaire on online teaching at Stockholm University during the Spring term 2021. Preliminary results using the Bayesian Regression Models using 'Stan’ (brms) R-Package show that students with special needs were adversely affected by the imposed online teaching. Among other things, the students felt less engaged intellectually, had lower confidence, felt examination instructions were unclear and spent more hours per week than their counterparts with no special needs.
- Seminar: Dan Hedlin, Department of Statistics, Stockholm University
Department of Statistics
A computed 95% confidence interval does cover the true value with probability 0.95 – epistemically interpreted Seminar Date: Wednesday 31 January 2024 Time: 13.00 – 14.00 Location: Campus Albano, Lecture room 9, house 2, level 2 Abstract How to interpret and make use of a computed (realised) confidence interval has been debated for a long time. I advocate that a realised confidence interval covers the parameter with probability . Suppose a sample of batteries in routine use is taken and their lifetimes are measured. A confidence interval is computed to days. The standard interpretation is that if we repeatedly draw samples of the same size and compute confidence intervals, 95% of the intervals will cover the unknown true life time. What can be said about the realised interval has not been clear in the literature. I shall give contradictory examples of interpretations. To appreciate the claim made here, we need to have a look at philosophy of probability and the Fisherian concept of relevant subsets.
- Seminar: Martin Hyllienmark, Department of Statistics, Stockholm University
Department of Statistics
Selection bias in big data in official statistics from a practitioner’s point of view Seminar Date: Wednesday 28 February 2024 Time: 13.00 – 14.00 Location: Campus Albano, Lecture room 15, house 2, level 2 Abstract Which is better: simple random sampling (SRS) with nonresponse or a large dataset with selection bias? Selection bias is increasingly problematic in surveys. Inspired by Meng’s (2018) paper Statistical paradises and paradoxes in big data, where he highlights statistical issues that bigness of data sets incur, we simulated sequences of growing populations with two different sampling designs: simple random sampling (SRS) with nonresponse and nonprobability data with selection bias (“big data”). The results showed that nonresponse combined with selection bias led to a constant bias and MSE only if the proportion of respondents was constant as the population size grew. Furthermore, there was a trade-off between bias and coverage probability caused by the amount of data available. By applying weights, it was possible to improve the results. The results in this study suggest that the data sources of choice may depend on the application of the statistics. Tools for comparing different data sources were investigated and discussed from a practical point of view.
- Seminar: Katarzyna Filipiak, Institute of Mathematics, Poznan University of Technology
Department of Statistics
Projection as the maximum likelihood estimator of the covariance structure Seminar Date: Wednesday 6 March 2024 Time: 13.00 – 14.00 Location: Campus Albano, Lecture room 27, house 4, level 2 Abstract Under the multivariate model with linearly structured covariance matrix with unknown variance components and known mean parameters it is known that the maximum likelihood estimators (MLEs) of variance components have explicit representation if and only if the space of covariance matrix is a quadratic subspace. In this talk, it will be shown, that the above result can be extended also for models with unknown mean parameters. Moreover, it will be proven that is such a model the projection of the MLE of unstructured covariance onto the space of structured matrices is still the MLEs as long as the estimated structure belongs to quadratic space.
- PhD thesis defence - Azadeh Chizarifard
Department of Statistics
On Estimation and Prediction in Linear Mixed Models. A new approach to studying equal BLUEs and BLUPs Thesis defence Date: Thursday 7 March 2024 Time: 10.00 – 13.00 Location: Campus Albano, auditorium 4, house 2, level 2 Abstract Linear mixed models (LMMs) are widely used to analyze repeated, longitudinal, or clustered data in many disciplines, such as biology, medicine, psychology, sociology, economics, etc. One of the essential components of a linear mixed model is its covariance structure, i.e., the covariance matrices of the random part and the error part, respectively, and the relationship between them. Specifying the covariance structure can be very challenging in a given situation. An important issue concerns the best linear unbiased estimators (BLUE) and the best linear unbiased predictors (BLUPs) under a given LMM and if two LMMs with different covariance structures generate equal BLUEs and/or BLUPs. This thesis proposes a new approach to determine the equality of the BLUEs and the BLUPs of the fixed and random effects under two linear mixed models with different covariance structures, and we present straightforward criteria to evaluate conditions that lead to equal BLUEs and BLUPs under these two models. Papers I and II focus on the equality of the BLUEs, while Papers II and IV study equal BLUPs under two LMMs with different covariance structures. More specifically, Paper I considers two different LMMs concerning their covariance matrices and whether the models generate equal BLUEs. In Paper II, we study the conditions for invariant BLUEs under a partitioned linear fixed effects model and a corresponding linear mixed model. Paper III develops the results to obtain the common BLUPs under two LMMs considering uncorrelated random effects and random errors. Finally, Paper IV extends the results of Paper III to obtain equal BLUPs in two linear mixed models, allowing for a correlation between the random effects and the random errors.
- Seminar: Daniel Buncic, Stockholm Business School
Department of Statistics
Ingress
- Seminar: Hannes Waldetoft, Dept. of Statistics, Uppsala University
Department of Statistics
Transformer assisted survey sampling for efficient finite population statistics in highly imbalanced textual data: public hate crime estimation Seminar Date: Wednesday 23 October 2024 Time: 13.00 – 14.00 Location: Campus Albano, lecture room 30, house 4, level 2 Abstract Estimating population parameters in finite populations of text documents can be challenging in cases where obtaining the labels (true values) for the target variable requires manual annotation. To address this problem, we combine predictions from a transformer encoder neural network with well-established survey sampling estimators. This is done by training a classifier and then using the model predictions as an auxiliary variable in the estimators. The applicability is demonstrated on Swedish hate crime statistics, which are based on Swedish police reports, for which approximately 1.5 million are being filed annually. Estimates of the yearly number of hate crimes and the police's under-reporting are derived using probability proportional to size (pps) sampling, regression estimation, and stratified random sampling. We conclude that if labeled training data is available, the proposed method can provide efficient estimates with reduced time spent on manual annotation.
- Seminar: Antonietta Mira, Professor of Statistics at Università della Svizzera italiana
Department of Statistics
Statistical Science meets Life Science Seminar Date: Friday 20 October 2023 Time: 14.00 – 15.00 Location: Theatre 3, Albano (ALB Hörsal 3, Albano Hus 2, Vån 3)
- Seminar: Martin Karlberg, Eurostat Luxembourg
Department of Statistics
Eurostat – the statistical office of the European Union Seminar Date: Friday 24 March 2023 Time: 11.30 – 12.30 Location: Campus Albano, lecture room 27, house 4, level 2 Abstract Eurostat is the statistical office of the European Union, and at the heart of the European Statistical System (ESS) – a collaboration with national statistical institutes which works to provide comparable statistics at the European level. Formally, Eurostat is a Directorate-General of the European Commission (one of the better-known European Union institutions). Martin Karlberg, who graduated from Stockholm University in the 1990s, joined Eurostat 20 years ago and is now deputy head of the Eurostat unit that is responsible for data & metadata services and standards. He will share his personal experience of working for Eurostat – and the European Commission in general (and how very different his current job often is from what he was trained to do during his statistics studies). His talk will cover various aspects of EU activities – in particular what Eurostat (and the ESS) actually does, and how the way of producing and disseminating European statistics (and official statistics in general) have changed (and may change) over time. The talk will also provide the audience with some information on what they could do to improve their chances to get the chance to work at or for Eurostat – or the European Union institutions in general.
- Seminar: Semhar Michael, South Dakota State University
Department of Statistics
Mixture Modeling of data with hierarchy Seminar Date: Wednesday 23 August 2023 Time: 13.00 – 14.00 Location: Campus Albano, Lecture room 25, house 4, level 2 Abstract Finite mixtures are known for modeling heterogeneity in data. The Gaussian mixture model is the most used by practitioners. The common way of estimating the parameters of this model assumes that the data is sampled through a simple random sampling process. However, in some applications such as the forensic source identification problem, data has a hierarchical structure in addition to the heterogeneity that occurs at different levels. In this work, we will discuss identifying and characterizing subpopulations when there are hierarchically structured data. This will be done through semi-supervised finite mixture models and by applying constraints that account for the hierarchy in the data. We will illustrate this based on a simulation study and a classical glass dataset. In addition, I will discuss some of the implications of this work on the forensic source identification problem.
- Seminar: Taras Bodnar, Department of Mathematics, Stockholm University
Department of Statistics
Practical and theoretical aspects of high-dimensional portfolio theory Seminar Date: Wednesday 19 April 2023 Time: 13.00 – 14.00 Location: Campus Albano, lecture room 29, house 4, level 2 Abstract Optimal asset allocation is considered in a high-dimensional asymptotic regime, namely when the number of assets and the sample size tend to infinity at the same rate. Due to the curse of dimensionality in the parameter estimation process, asset allocation for such portfolios becomes a challenging task. Using techniques from random matrix theory, we derive shrinkage estimators for the weights of the mean-variance optimal portfolios. Also, new inferential procedures based on the optimal shrinkage intensity for testing the mean-variance efficiency of a high-dimensional portfolio are developed and the asymptotic distributions of the proposed test statistics are derived. In extensive simulations, we show that the suggested tests have excellent performance characteristics for various values of concentration ratio. The practical advantage of the proposed procedures is demonstrated in an empirical study based on stocks included into the S&P 500 index. We found that there are periods of time where one can clearly reject the null hypothesis of mean-variance efficiency of the equally weighted portfolio. Moreover, the mean-variance portfolio outperforms the equally weighted portfolio in these periods.
- Seminar: Johan Koskinen, Department of Statistics, Stockholm University
Department of Statistics
Analysing network data that are partially observed Seminar Date: Wednesday 8 February 2023 Time: 13.00 – 14.00 Location: Campus Albano, classroom 26, house 4, floor 2 Abstract Network data may be represented as binary graphs, either directed or undirected, and have a long history of being used to model and describe interaction between people and other entities, with formal approaches dating back to at least the start of the twentieth century. For a graph, the potential tie between a pair of nodes is represented by a binary indicator variable that we may call a tie-variable. These tie-variables are indexed by the labels of the nodes and can be organised in a so-called node by node adjacency matrix. Since the entries of the adjacency matrix are cross-classified by both the row node, and the column node, the tie-variables are highly interdependent. Exponential (family) random graph models (ERGMs) constitute a class of log-linear models with natural parameters that have as statistics a subset of graph statistics derived out of principled dependence assumptions. Due to these dependencies, the ERGM for a network does not marginalise and subgraphs of the network do not follow models of the same form. Here we discuss inference approaches for the parameters of the ERGM when some tie-variables are missing. The treatment of missing data in ERGM also applies to cases where data are missing by design, for example when the network data have been obtained through a link-tracing designs, such as snowball sampling. We describe a Bayesian approach for estimation and provide examples of applications to networks of young men who have sex with men, rebels in the Democratic Republic of the Congo, as well as the use of the Bayesian estimation scheme for imputing initial conditions in the analysis of network panel data. The latter case is illustrated with an application to social support networks in bushfire-affected communities in Australia. The use of the proposed approach is contingent of a number of fairly heroic assumptions, some of which will be brought up for discussion.
- Today´s Seminar with Sebastian Ankargren, Spotify, has been cancelled
Department of Statistics
Randomized controlled trials at scale: A/B testing at Spotify Seminar Date: Wednesday 25 January 2023 Time: 13.00 – 14.00 Location: Campus Albano, classroom 27, house 4, floor 2 Abstract Abstract: At Spotify, we have hundreds of experiments, so-called A/B tests, running at any given point in time. A/B testing is a form of online controlled experimentation that builds on the foundational ideas of randomized controlled trials (RCTs). What makes A/B tests different from RCTs is the environment in which they are used and the requirements they face. A/B tests are often cheap, they can easily involve hundreds of millions of units, they can often have effects that are heterogeneous across units, and they can be risky but only in terms of degrading the user experience. In this talk, I will give a high-level overview of how we do online experimentation at Spotify and present two challenges we currently have: sequential inference when outcomes are measured over open-ended windows of time, and finding groups of users with heterogeneous treatment effects.
- Dissertation defence - Renata Tsirpitzi
Department of Statistics
Due to new advice and recommendations to limit the spread of covid-19: Before entering the auditorium, we recommend that you wash your hands with soap and water or use hand disinfectant. Inside the auditorium, do not sit too close to each other in larger gatherings / groups. Dissertation Date: Friday 17 December 2021 Time: 13.00 – 16.00 Location: Hörsal 11, Frescati. Zoom: https://stockholmuniversity.zoom.us/j/69323772943 Optimal design for dose-finding studies Abstract One of the most complex tasks during the clinical development of a new drug is to find a correct dose. Optimal experimental design has as a goal to find the best ways to perform an experiment considering the available resources and the statistical model. Optimal designs have already been used to determine the design of dose-finding studies. In this thesis, optimal designs are considered for the simultaneous response of efficacy and safety in a bivariate model, for the drug combination trials, and for general regression problems, including but not limited to dose-finding analysis. The thesis consists of four papers: In Paper I, the dose that maximizes the clinical utility index based on an efficacy-safety Emax model gives us the desirable balance between effects and side effects. In order to make use of a symmetry property, we use a log-transformed dose scale. The geometric characterization of the multivariate Elfving method is used to derive c-optimal points and weights for arbitrary c-vectors. The second paper is an extension of the first one. We still use the log-transformed dose scale bivariate model and consider now also the placebo effect and side-effect. Fedorov’s exchange algorithm is applied in order to derive locally D-optimal designs numerically.
- Dissertation defence - Jonas Bjermo
Department of Statistics
Due to new advice and recommendations to limit the spread of covid-19: Before entering the auditorium, we recommend that you wash your hands with soap and water or use hand disinfectant. Inside the auditorium, do not sit too close to each other in larger gatherings / groups. Dissertation Date: Friday 10 December 2021 Time: 10.00 – 13.00 Location: Hörsal 4, hus 2,Aalbano. Zoom: https://stockholmuniversity.zoom.us/j/69323772943 Test design for mean ability growth and optimal item calibration for achievement tests Abstract In this thesis, we examine two topics in the area of educational measurement. The first topic studies how to best design two achievement tests with common items such that a population mean-ability growth is measured as precisely as possible. The second examines how to calibrate newly developed test items optimally. These topics are two optimal design problems in achievement testing. Paper I consist of a simulation study where different item difficulty allocations are compared regarding the precision of mean ability growth when controlling for estimation method and item difficulty span. We take a more theoretical approach on how to allocate the item difficulties in Paper II. We use particle swarm optimization on a multi-objective weighted sum to determine an exact design of the two tests with common items. The outcome relies on asymptotic results of the test information function. The general conclusion of both papers is that we should allocate the common items in the middle of the difficulty span, with the two separate test items on different sides. When we decrease the difference in mean ability between the groups, the ranges of the common and test items coincide more. In the second part, we examine how to apply an existing optimal calibration method and algorithm using data from the Swedish Scholastic Aptitude Test (SweSAT). We further develop it to consider uncertainty in the examinees' ability estimates. Paper III compares the optimal calibration method with random allocation of items to examinees in a simulation study using different measures. In most cases, the optimal design method estimates the calibration items more efficiently. Also, we can identify for what kind of items the method works worse. The method applied in Paper III assumes that the estimated abilities are the true ones. In Paper IV, we further develop the method to handle uncertainty in the ability estimates which are based on an operational test. We examine the asymptotic result and compare it to the case of known abilities. The optimal design using estimates approaches the optimal design assuming true abilities for increasing information from the operational test.
- Seminar: Mathias Millberg Lindholm, Dept. of Mathematics, Stockholm University
Department of Statistics
Tree-based boosting and varying coefficient models Seminar Date: Wednesday 6 November 2024 Time: 13.00 – 14.00 Location: Campus Albano, lecture room 23, house 4, level 2 Abstract Tree-based boosting models, such as gradient boosting machines (GBMs), have proved to have good predictive performance, at the same time being simple to use for practitioners. Current off-the-shelf GBM program-packages provide natural alternatives to, e.g., GLMs and GAMs. These implementations, and the original GBM-formulation, focus on univariate models, modelling an unknown mean function. In this talk I will discuss how to implement and tune cyclic multidimensional GBMs (CGBMs) that can be used for distributional regression, where you estimate an unknown multidimensional parameter function depending on covariates. That is, CGBMs can be used to jointly model, e.g., the shape and rate parameters in a Gamma-distribution as functions of covariates. A different type of application of the CGBM is to use it to construct more interpretable univariate models: By using the CGBM we introduce an univariate tree-based varying coefficient model, which combines the predictive performance of GBMs with an interpretable GLM-like structure.
- Seminar: Henrik Imberg, Univeristy of Gothenburg
Department of Statistics
Active sampling: A machine-learning-assisted framework for finite population inference with optimal subsamples Seminar Date: Wednesday 18 December 2024 Time: 13.00 – 14.00 Location: Campus Albano, lecture room 29, house 4, level 2 Abstract Data subsampling has become widely recognized as a tool to overcome computational and economic bottlenecks in analyzing massive datasets. We contribute to the development of adaptive design for estimation of finite population characteristics, using active learning and adaptive importance sampling. We propose an active sampling strategy that iterates between estimation and data collection with optimal subsamples, guided by machine learning predictions on yet unseen data. The method is illustrated on virtual simulation-based safety assessment of advanced driver assistance systems. Substantial performance improvements are demonstrated compared to traditional sampling methods. The presentation is based on our recent Technometrics paper, published in 2024, available at: https://doi.org/10.1080/00401706.2024.2374554
- Seminar: Yukai Yang, Dept. of Statistics, Uppsala University
Department of Statistics
Moment matching based sparse and noisy point cloud registration Seminar Date: Wednesday 12 February 2025 Time: 13.00 – 14.00 Location: Campus Albano, lecture room 26, house 4, level 2 Abstract Point cloud registration is a fundamental problem in Simultaneous Localization and Mapping (SLAM), particularly in robotics applications where sensor data is often sparse and noisy. In this paper, we introduce a novel approach to point cloud registration based on moment matching, specifically designed to handle challenging conditions such as low point density and measurement noise. Our method is applied to millimeter-wave radar data collected from a moving vehicle, demonstrating its effectiveness in real-world scenarios. A key advantage of our approach is its computational efficiency. We ensure that registration between consecutive frames is completed within 100 milliseconds, making it suitable for real-time applications. Additionally, we evaluate the accuracy of our method by comparing it against widely used registration techniques. Despite operating on significantly sparser radar data, our method achieves comparable accuracy to state-of-the-art LiDAR-based approaches, which rely on much denser point clouds. These results highlight the potential of moment matching as an efficient and robust solution for point cloud registration in SLAM, particularly in resource-constrained environments where dense LiDAR scans are not available.