Date : Vendredi 24 janvier 2025
Heure : 10h30
Salle : 6214, pavillon André-Aisenstadt

Conférencière: Cindy Xinyi Zhang, postdoctoral fellow at Johns Hopkins University,

Titre : On Statistical Solutions for Complex Data Challenges.

Résumé : In this talk, I will present two lines of research: causal inference and statistical neuroimaging.

The first part focuses on causal discovery from observational studies using instrumental variables to address unmeasured confounding. A key challenge in applying instrumental variable methods is identifying valid instruments from a potentially large candidate set. In practice, many candidate instruments are irrelevant to the exposure of interest. Additionally, some relevant candidates may have direct effects on the outcome, introducing bias into causal effect estimation and making them invalid. I will first discuss the challenges posed by irrelevant variables that exhibit spurious correlations with the exposure. Then, I will present a data-driven approach that identifies valid instruments from a large candidate pool by addressing both the issues of irrelevant variables and invalid instruments. This method leverages “pseudo” variables constructed to mimic irrelevant candidates as a key component.

In the second part, I will discuss a multiple testing procedure developed to balance family-wise error rate control and statistical power. This work is motivated by the multiplicity issue inherent in functional magnetic resonance imaging (fMRI) studies, where the large number of comparisons necessitates robust statistical adjustments.