Causal Inference using Bayesian non-parametric quasi-experimental designs for sequential data

Abstract

In many domains, such as healthcare, it is crucial to know as fast as possible whether an intervention (e.g. drug treatment) had an e ect. In these contexts, time is of the essence and samples are both costly to acquire and limited in their amount, as they only arrive sequentially over time. This research proposes a method called Sequential Bayesian non-parametric quasi-experimental designs (S-BNQD) which allows to establish a causal e ect in sequential settings. S-BNQD is based on the Bayesian non-parametric quasi-experimental design (BNQD) (Hinne, van Gerven, & Ambrogioni, 2019) which is a Bayesian version of quasi-experimental designs. BNQD is based on Gaussian Process Regression which makes it particularly attractive for sequential designs due to its exi- bility and uncertainty estimation (M. M. Zhang, Dumitrascu, Williamson, Engelhardt, & Oct, 2019). The new framework can be applied in various contexts where Randomized Control Trials (RCT's) are not practical and it is crucial to know as reliably and quickly as possible whether an intervention was e ective or not.