Title:Automated Meta-Analysis: A Causal Learning Perspective

Abstract:Meta-analysis is a systematic approach for understanding a phenomenon by analyzing the results of many previously published experimental studies. It is central to deriving conclusions about the summary effect of treatments and interventions in medicine, poverty alleviation, and other applications with social impact. Unfortunately, meta-analysis involves great human effort, rendering a process that is extremely inefficient and vulnerable to human bias. To overcome these issues, we work toward automating meta-analysis with a focus on controlling for risks of bias. In particular, we first extract information from scientific publications written in natural language. From a novel causal learning perspective, we then propose to frame automated meta-analysis -- based on the input of the first step -- as a multiple-causal-inference problem where the summary effect is obtained through intervention. Built upon existing efforts for automating the initial steps of meta-analysis, the proposed approach achieves the goal of automated meta-analysis and largely reduces the human effort involved. Evaluations on synthetic and semi-synthetic datasets show that this approach can yield promising results.