Variational autoencoders and deep neural networks

MiSoC PhD student Damian Machlanksi, and Co-Is Spyros Samothrakis and Paul Clarke are leading a new project that involves the use of variational autoencoders and deep neural networks for causal effect estimation and causal discovery.

Focusing on variational autoencoders (VARs), these are a form of generative model and so a form of unsupervised ML. These models do not simply predict outcomes from predictors but learn complex mixtures of normal distributions to accurately approximate the joint probability distribution of the observed data.

Other forms of generative models from ML are now being used in causal econometrics to simulate potential outcomes from models which match the complexity of real probability distributions so that econometricians can carry out simulation experiments to assess the performance of estimators under realistic conditions (Athey et al. 2021).

We are initially focussing instead on using VARs (Kingma and Welling 2014) to learn complex causal models for X, T and Y from sample data and use the results to estimate both average treatment effects and heterogeneous treatment effects (referred to as conditional average treatment effects). This novel work is demonstrates that this approach can match the performance of specialist ML learners for causal effects (e.g. Guo et al. 2020) when we impose a graph on the data that represents the assumptions the analyst is prepared to make.

This approach adopts a far more general framework than that used by the specialized ML methods, one that will allow us to focus on applying ideas from the literature on causal discovery (e.g. Peters et al. 2017) to social science data. In theory, it is possible to learn the graph representing the temporal and causal ordering between variables (thus allowing one to correctly identify whether a variable can be included in X or not) and, given this graph, estimate the causal effect. Within the no-unobserved-confounding framework, this would allow analysts to road test the assumptions they needed to estimate causal effects by viewing the causal graph discovered by the learner.

Our initial findings suggest that identifying the true causal graph is beyond the capability of these methods. But we are investigating whether it is possible a) to partially identify – that is, identify the set of – alternative graphs consistent with the observed data, and b) to use reinforcement learning to instruct analysts about the additional data they need to collect to learn more accurately what the true graph is.