講演者：Adriaan Ludl 先生（University of Bergen）
題目： Causal Inference for Time Series and Population Data
There are a variety of causal inference methods that provide estimates of causal relationships given observational data. In this talk, I will present methods for time series data and for population data which can be used to obtain regulatory networks.
The first part will focus on methods that can be applied to time series. One such method is transfer entropy (TE) which is an asymmetric and non-linear measure of the relationship between two variables. This method can be applied to data collected from living neuronal cultures in vitro. Our evaluation on simulations of neuronal activity shows that it performs well.
In the second part, we will consider methods which are suitable for population data. We will focus on Bayesian methods that are based on mediation and instrumental variables. These methods can be applied to genomic data and can help to understand the regulatory relationships between genes. Our evaluation of these methods on a data set of a yeast cross shows that these methods can play complementary roles for identifying causal genes. The performance of mediation saturates at large sample sizes, while instrumental variable methods tend to make false positive predictions due to genomic linkage of the instruments.
We are currently developing a method to estimate the strength of causal interactions in cases where multiple variables share multiple instruments. Our first evaluation on regulatory hotspots in yeast shows that this method can address the shortcomings above.
Finally, we will consider some of the general limitations and pitfalls of causal inference.
(ホスト: 清水昌平先生 1時間の予定ですが多少前後するかもしれません)