The dilemma of establishing causal relationships in areas such as medicine, economics, and social sciences is characterized as the “Fundamental Problem of Causal Inference”. When observing an outcome, it is often unclear what the result might have been under a different intervention. Various indirect methods have been developed to estimate causal effects from observational data to address this concern.
Existing methods include the S-Learner, which uses a single model with the treatment variable as a feature, and the T-Learner, which uses separate models for treated and untreated groups. However, these approaches have their own set of problems. The S-Learner often has a bias towards zero treatment effect, while the T-Learner can be data inefficient.
More advanced models such as TARNet, Dragonnet, and BCAUSS have been developed, which utilize the principle of representation learning with neural networks. These neural networks compose of a pre-representation component that extracts representations from input data, and a post-representation component that maps these representations to the intended output.
Regardless of promising results from these representation-based methods, they tend to overlook a specific source of bias: spurious interactions between variables within the model. Such interactions can skew estimated causal effects when data is limited.
In order to tackle this issue, researchers from the Universitat de Barcelona have introduced a new technique known as Neural Networks with Causal Graph Constraints (NN-CGC). The basic idea of NN-CGC is to limit the learned distribution of the neural network to better uphold the causal model, efficiently reducing dependence on spurious interactions.
The method with which NN-CGC operates is simple: first, the input variables are divided into groups based upon the causal graph available. Second, each variable group is processed independently through a series of layers which model the Independent Causal Mechanisms for the outcome variable and its direct causes. Third, NN-CGC ensures that the learned representations are free from spurious interactions by processing each variable group separately. Finally, the outputs from the independent group representations are combined, and transferred through a linear layer to form the final representation.
The researchers subjected NN-CGC to various synthetic and semi-synthetic benchmarks, including the widely-recognized IHDP and JOBS datasets. The findings indicate that the constrained versions of TARNet, Dragonnet, and BCAUSS (together with NN-CGC) consistently beat their unconstrained competitors.
In conclusion, NN-CGC offers a new and flexible technique for integrating causal information into neural networks for causal effect estimation. It manages to address the often-neglected issue of spurious interactions, showing significant improvements over existing methods. The researchers have made this a public asset, allowing others to build upon and enhance this promising technique.