Kolyan Ray - Semiparametric Bayesian Causal Inference and Uncertainty Quantification
| Dates: | 19 October 2022 |
|---|
| Times: | 15:00 - 16:00 |
|---|
| What is it: | Seminar |
|---|
| Organiser: | Department of Mathematics |
|---|
| Who is it for: | University staff, External researchers, Current University students |
|---|
| Speaker: | Dr Koylan Ray |
|---|
|
Join us for this research seminar, part of the SQUIDS (Statistics, quantification of uncertainty, inverse problems and data science) seminar series.
Abstract: Inferring the causal effect of a treatment or condition based on observational data (i.e. not a randomized controlled trial) is an important problem in many applications. This is a challenging problem due to the selection bias and missing counterfactuals. We investigate Bayesian inference for (population) average treatment effects in such a setting.
We show that standard Gaussian priors can yield asymptotically optimal inference for this task, including uncertainty quantification, for sufficiently regular true functions. However, these conditions become harder to satisfy as the dimension of the covariate space increases, meaning such priors can yield badly biased inference in high dimensions. We propose a propensity score-based prior modification that corrects for the first-order posterior bias, leading to efficient inference under weaker conditions. Numerical simulations confirm significant improvement in both estimation accuracy and uncertainty quantification with this modification.
Speaker
Dr Koylan Ray
Organisation: Imperial
Travel and Contact Information
Find event
Frank Adams 2
Alan Turing Building
Manchester
