Deep learning framework for semiparametric causal mediation analysis
This technology is a deep-learning-based causal mediation analysis tool that allows estimation of natural direct and indirect causal effects without imposing sparsity constraints on neural-network architecture.
Unmet Need: Accurate bias-robust mediation analysis for complex data
Causal mediation is routinely used to disentangle causal pathways in clinical trials, social science studies, and fairness audits. Current approaches for causal mediation are largely based on rigid parametric models or deep-learning-based estimators whose statistical guarantees fail in high-dimensional settings and suffer from bias. Moreover, existing methods often lack validation on large-scale data with high-dimensional confounding, leaving a gap between practice and theory.
The Technology: Bias-reduced mediation via non-sparse deep-neural nets
This technology, called DeepMed, is a deep-neural-network (DNN) based approach to calculate more precise nuisance function estimates and more precise causal effect estimates. By formulating estimation around the multiply-robust influence function, the procedure remains consistent if at least one of the models is correctly specified. The results from simulated data show that it reduces bias in estimating natural direct and indirect effects, and works well with dense neural networks as well. The technology comes with an R package with a user-friendly interface.
Applications:
- Algorithmic-fairness auditing
- Clinical and biomedical research
- Policy evaluation
- Industrial A/B testing
Advantages:
- Semiparametric optimality with less bias
- No sparsity requirement for neural net architecture
- Theoretical guarantees with fewer assumptions
- Open-source R package
Lead Inventor:
Related Publications:
Tech Ventures Reference:
IR CU25372
Licensing Contact: Joan Martinez