# Joint Causal Inference on Observational and Experimental Datasets

**Authors:** Sara Magliacane, Tom Claassen, Joris M. Mooij  
**arXiv:** [1611.10351](https://arxiv.org/abs/1611.10351)  
**Date:** 2016-11-30  
**Summary page:** https://www.summarizepaper.com/en/paper/1611.10351v1/

## Key points

- Joint Causal Inference (JCI) is a powerful approach to causal discovery that operates over multiple datasets.
- JCI enables simultaneous learning of causal structure and intervention targets by leveraging statistical independences in pooled data.
- JCI allows for various types of interventions, providing flexibility in exploring different causal relationships.
- JCI can learn intervention targets, which is crucial for understanding how specific variables are affected by interventions.
- JCI systematically pools data across different datasets, enhancing the statistical power of independence tests and improving the accuracy and identifiability of predicted causal relations.
- Faithfulness violations due to deterministic relations are a technical challenge in JCI.
- ACID (Ancestral Causal Inference with Determinism) is proposed as a strategy within JCI to address faithfulness violations.
- ACID leverages redundant information in the data to improve output reliability.
- The benefits of JCI with ACID are demonstrated using a simulated dataset, showcasing how it enhances causal discovery from observational and experimental datasets.
- Joint Causal Inference is presented as a robust framework for discovering causality across multiple datasets.

## Summary

In their paper titled "Joint Causal Inference on Observational and Experimental Datasets," authors Sara Magliacane, Tom Claassen, and Joris M. Mooij introduce Joint Causal Inference (JCI), a powerful approach to causal discovery that operates over multiple datasets. JCI enables the simultaneous learning of both the causal structure and targets of interventions by leveraging statistical independences in pooled data. Compared to existing constraint-based methods for causal discovery from multiple datasets, JCI offers several advantages. Firstly, it allows for various types of interventions, providing flexibility in exploring different causal relationships. Additionally, JCI can learn intervention targets, which is crucial for understanding how specific variables are affected by interventions. Moreover, JCI systematically pools data across different datasets, enhancing the statistical power of independence tests and improving the accuracy and identifiability of predicted causal relations. However, one technical challenge that arises in JCI is the occurrence of faithfulness violations due to deterministic relations. To address this issue, the authors propose a simple yet effective strategy within ACID (Ancestral Causal Inference with Determinism), an extension of Ancestral Causal Inference (ACI). ACID is a logic-based method for causal discovery that leverages redundant information in the data to improve output reliability. To demonstrate the benefits of JCI with ACID, the authors evaluate their approach using a simulated dataset. This evaluation showcases how JCI enhances causal discovery by jointly inferring causal structures and intervention targets from observational and experimental datasets. Overall, this paper presents Joint Causal Inference as a robust framework for discovering causality across multiple datasets. By addressing faithfulness violations through ACID's determinism-tolerant strategy, JCI improves upon existing approaches while offering increased flexibility and accuracy in identifying causal relationships and intervention targets.