日時：2026年2月27日(金) 10:00 ~　90分

場所：webのみ　（※ 対象は特に限定しない）

講演者：Mariyam Khan 先生（University of Bergen, Norway）　

題目： Causal Inference with Hidden Confounding: Negative Controls, Proximal Causal Learning, and Tetrad-Based Discovery

概要：
Unmeasured confounding is a central obstacle to causal
inference in observational studies and remains a key limitation in
practice, including in Mendelian Randomization when assumptions are
threatened by mechanisms such as population stratification. In this
talk, I will introduce negative controls, variables that are affected
by an unmeasured confounder but are not causally affected by the
treatment or the outcome of interest. Negative controls can reveal
hidden confounding bias and, under suitable graphical structure, can
be used as proxy variables to identify causal effects. I will then
describe Proximal Causal Learning (PCL), which uses pairs of such
negative controls as proxies for latent confounders to identify causal
effects, recovering the target causal estimand under unmeasured
confounding.

The second part focuses on a major practical bottleneck: negative
controls are typically chosen using domain knowledge, and their
exclusion restrictions are difficult to justify and validate in
high-dimensional settings such as genetic studies with thousands of
candidate proxies. I will present a data-driven causal discovery
approach that screens candidate genes to identify valid disconnected
negative control triplets which are sets of three variables that share
a single latent confounder but have no direct connections among
themselves, or to the treatment and outcome, beyond this confounder.
This approach exploits rank and tetrad constraints implied by such
one-factor latent structures. In particular, the key testable
implication is a vanishing tetrad, equivalent to a zero determinant of
a 2×2 subcovariance block.

I will discuss statistical validation via classical tetrad tests and
address the multiple-testing challenges induced by scanning many
candidate triplets. Critically, triplets passing these tests can then
be used as proxies in PCL to recover causal effects, with false
discovery rate (FVR) control ensuring that the bias from invalid
selections remains bounded. This approach enables scalable
confounder-robust causal inference in settings where traditional
methods fail.

お問い合わせは以下のアドレスにメールでご連絡ください。
dser-center@biwako.shiga-u.ac.jp