Abstract: Meta-analysis is a systematic approach for understanding a phenomenon by analyzing the results of many previously published experimental studies related to the same treatment and outcome measurement. It is an important tool for medical researchers and clinicians to derive reliable conclusions regarding the overall effect of treatments and interventions (e.g., drugs) on a certain outcome (e.g., the severity of a disease). Unfortunately, conventional meta-analysis involves great human effort, i.e., it is constructed by hand and is extremely time-consuming and labor-intensive, rendering a process that is inefficient in practice and vulnerable to human bias. To overcome these challenges, we work toward automating meta-analysis with a focus on controlling for the potential biases. Automating meta-analysis consists of two major steps: (1) extracting information from scientific publications written in natural language, which is different and noisier than what humans typically extract when conducting a meta-analysis; and (2) modeling meta-analysis, from a novel \textit{causal-inference} perspective, to control for the potential biases and summarize the treatment effect from the outputs of the first step. Since sufficient prior work exists for the first step, this study focuses on the second step. The core contribution of this work is a multiple causal inference algorithm tailored to the potentially noisy and biased information automatically extracted by current natural language processing systems. Empirical evaluations on both synthetic and semi-synthetic data show that the proposed approach for automated meta-analysis yields high-quality performance.