The performance of prediction models and the estimated effects of risk factors in health data science depends strongly on how the (test) data is sampled, and whether it is representative of the population that we aim to describe. If the dataset is obtained in a non-random manner, e.g. via invitation to participate, it can result in a selection bias. and confound both risk factor estimation and prediction model accuracy.

There exist many types of selection biases (e.g. participation bias, survival bias, ascertainment bias, immortal time bias, and collider bias) and they can severely confound both risk factor estimation and prediction model accuracy. In data science, unaddressed selection bias cannot only lead to biased external validation results but also impact model generalizability, reproducibility, fairness, and downstream decision-making.

This session brings together researchers with diverse expertise to discuss methodological approaches for detecting, quantifying, and correcting for selection bias with a focus on health science use cases.

Program

A 90 minutes seminar with 3 speakers (each allotted 20 minutes), and a 30 minutes interactive panel discussion at the end.

What is selection bias and why study design matters (20 minutes)
Theresa Wimberley Böttger, Senior Researcher, Aarhus University

Mitigating selection bias: Generalizing evidence from clinical trials to target populations (20 minutes)
Zehao Su, Postdoc, University of Copenhagen

Using genetics to test models assumptions in time-to-event models and detect biases (20 minutes)
Genona T. Maseras, Research Assistant, Aarhus University

Panel discussion and Q&A (30 minutes)
Speakers + Invited panel member Prof. Esben Agerbo, Aarhus University + Moderator Assist. Prof. Anne Helby Petersen, University of Copenhagen

Organizers

• Research Assist. Genona T. Maseras, Aarhus University
• Assist. Prof. Adrian Zucco, University of Copenhagen
• Assist. Prof. Anne Helby Petersen, University of Copenhagen
• Prof. Bjarni Vilhjalmsson, Aarhus University

Level

Introductory: Suitable for beginners with little to no prior knowledge of the subject.