Co-Developing Causal Graphs with Domain Experts Guided by Weighted FDR-Adjusted p-values

Expert-led causal graph design improves AI responsibility: The paper demonstrates that having domain experts actively co-design causal graphs with statisticians—rather than letting algorithms determine relationships automatically—helps ensure AI systems are grounded in real-world knowledge and human oversight, which is essential for responsible AI practices.

Statistical feedback makes expert intuition more rigorous: By overlaying False Discovery Rate (FDR)-adjusted p-values onto causal graphs, the approach gives domain experts concrete statistical evidence about which proposed causal relationships are supported by data, helping distinguish between expert beliefs that are well-founded versus those needing reconsideration.

The method addresses multiple types of causal questions simultaneously: The framework evaluates correlation strengths, directions of causal effects, and whether statistical models actually match observed data patterns—providing comprehensive feedback rather than single-point estimates, which reduces the risk of oversimplified or misleading causal conclusions.

Iterative refinement reduces bias in causal assumptions: The paper describes a cyclical process where initial expert-drafted graphs are tested against data, then refined based on statistical results, preventing the common problem where either pure expert opinion or pure statistical analysis alone leads to flawed causal models that could drive poor decisions.

Incorporating expert rankings into statistical methods personalizes rigor: The approach allows experts to rank which causal relationships they're most confident about, and these rankings can be incorporated into the FDCR control method, creating a principled way to balance domain knowledge with statistical evidence rather than treating them as competing approaches.