Providing robust causal analyses for business use cases

• Data & AI - Strategy & Consulting Dept., Accenture, Italy [francesca.gerardi@accenture.com]

1 Abstract

Causal inference is the process of determining whether two variables are connected through a cause-effect relationship, allowing us to go beyond standard statistical analysis and perform true root-cause investigation. Understanding these causal relationships enables not only the estimation of variables impacts, but also the simulation of hypothetical or future scenarios, entering the realm of interventions and, ultimately, counterfactuals [Pearl, 2009, Pearl and Mackenzie, 2018]. The latter includes questions such as “What would have happened if…’, which are the core of this class of algorithms.

This talk introduces the application of causal inference to real-world business cases, with a particular focus on the marketing and retail area. The causal approach can indeed be extremely valuable towards the definition of data-driven strategies aimed at the growth of a business. However, starting from the assumption of a causal graph, the reliability of causal inference is challenged by unknown confounders and biases, not to mention the issues related to real-world data imperfection.

In order to enhance the robustness of the graph and hence of the analysis, we propose a framework based on the work of Eulig et al. [2024], combining careful modeling with falsification strategies to assess whether a given causal graph offers statistically significant explanatory power. By iteratively refining models through falsification and applying causal minimality suggestions, we move toward more trustworthy, statistically robust, and actionable conclusions.