Data Analysis and Statistical Science

Quetelet seminars 2018

Quetelet seminars 2018

      Wen Wei Loh

      28 May 2018

 

ABSTRACT LIST

Why I love statistics

Murthy N Mittinty

University of Adelaide, Australia

Wednesday, May 9, 2018, 12h30

Auditorium A3, Campus Sterre, S9, Krijgslaan 281, 9000 Gent

Abstract:

“Statistics is a field that lets a thousand flowers bloom” (Amanda Gobeck). To fully understand the power of statistics we need four fundamental aspects; an open mind, an eagerness to listen, a desire for being honest, and admitting that there are multiple pathways to attain truth.  These are like the four fundamental operators, +, -, / and * which every algebra requires. Keeping an open mind allows us to not fall prey to selectivism, restricting to a school of statistics. Moreover, it also allows one to be creative. Listening allows us to see that the answer is in the problem. Being honest allows us to be, transparent with our code, model and data, open. Acknowledging that there are multiple pathways to truth allows us to appreciate the power of uncertainty, speak upfront the abilities and inability of what we can and cannot do with the data on hand. In this presentation I would like to share how I fell in love with statistics by chance and decided to stay by choice.

Identifying treatment responders using counterfactual modeling and potential outcomes

Raphaël Porcher

Paris Descartes University, France

Thursday, May 17, 12h30

room 3.1, Campus Sterre, S9, Krijgslaan 281, 9000 Gent

Abstract:

Individualizing treatment according to patients’ characteristics is central for personalized or precision medicine. There has been considerable recent research in developing statistical methods to determine optimal personalized treatment strategies by modeling the outcome of patients according to relevant covariates under each of the alternative treatments, and then relying on so-called predicted individual treatment effects. In this paper we use potential outcomes and principal stratification frameworks to develop a multinomial model for left and right-censored data to estimate the probability a patient being a responder given a set of baseline covariates in the context of RCT but also in observational study. This method is based on the monotonicity assumption, meaning that no patients would respond to the control treatment while not responding to the experimental one. We conduct a simulation study to evaluate the properties of the proposed estimation method. Results have shown that the predictions of the probability of being a responder were well calibrated even if we observed variability and a negligible bias. We finally applied the method on a cohort study on the selection of patients for additional radiotherapy after resection of a soft-tissue sarcoma.

Causal Mediation Analysis for Randomised Studies with Longitudinal Data using Structural Equation Modelling

Wen Wei Loh

Ghent University, Belgium

Monday, May 28, 12h30

Auditorium 3, Campus Dunant, Henri Dunantlaan, 9000 Gent

Abstract:

In a randomised study with longitudinal data on the mediator and outcome, the direct effect of the treatment on the outcome at a particular time includes all pathways that avoid earlier instances of the mediator. Estimation of the direct effect thus requires adjusting for confounders between the outcome and earlier instances of the mediator. But when the set of confounders are themselves affected by treatment, standard regression adjustment is prone to possibly severe bias. Under a certain class of linear models, traditional path analysis methods provide unbiased estimates of the controlled direct effect, which are obtained by combining the estimated path coefficients for the constituent paths. We describe how the path analysis approach can be embedded within the structural equation modelling framework and propose extensions to settings with latent mediator and outcome. However, when time-varying exposure-confounder and mediator-confounder interactions are present, the path analysis approach can produce biased estimates. We propose a G-estimation approach that can incorporate nonlinearities and interactions in the presence of post-treatment confounding and yields unbiased estimates of the controlled direct effects. The G-estimation approach adapts existing methods for time-varying treatments and generalizes the proposed path analysis approach to settings with noncontinuous mediators and confounders.

This is joint work with Beatrijs Moerkerke, Tom Loeys and Stijn Vansteelandt.

A deep but sweet introduction to depth

Germain Van Bever
Mathematics Department and ECARES Université libre de Bruxelles, Belgium

Wednesday, January 24, 2018, 13h
room 3.1, Campus Sterre, S9, Krijgslaan 281, 9000 Gent

Abstract:

This talk aims to be a non-exhaustive and subjective review of the field of statistical depth.
The introduction will cover the concept of statistical depth and the center-outward, quantile, ordering of the observations it provides. Famous examples, such as the halfspace depth (Tukey, 1975) and simplicial depth (Liu, 1990) will be described. The rest of the talk will cover two topics in more details: depth-based techniques for classification and depth functions outside the location setup.
In the context of classification, a new class of depth-based classification procedures that are of a nearest neighbor nature will be presented. The corresponding nearest neighbors are identified through a symmetrized depth construction. The classifiers enjoy (i) the robustness and affine-invariance of depth-based procedures and (ii) the good asymptotic properties of nearest-neighbor classifiers. In particular, the proposed depth-based classifiers are universally $L^p$-consistent.
Halfspace depth concepts for scatter, concentration and shape matrices will also be proposed. Interestingly, fully understanding these depths requires considering different geometries/topologies on the space of scatter matrices. In the spirit of Zuo and Serfling (2000), the structural properties a scatter depth should satisfy, and whether or not these are met by the proposed depth, will be investigated.
Throughout the talk, we illustrate the practical relevance of the proposed concepts by considering simulated and real data examples.