Causal inference using global forecasting models for counterfactual prediction

This research proposes a global forecasting and inference method based on recurrent neural networks (RNN) to predict policy interventions’ causal effects on an outcome over time through the counterfactual approach. The traditional univariate methods that operate within the well-established synthetic control method have strong linearity assumptions in the covariates. This has recently been addressed by successfully using univariate RNNs for this task. We use an RNN trained not univariately per series but globally across all time series, which allows us to model treated and control time series simultaneously over the pre-treatment period. Therewith, we do not need to make equivalence assumptions between distributions of the control and treated outcomes in the pre-treatment period. This allows us to achieve better accuracy and precisely isolate the effect of an intervention. We compare our novel approach with local univariate approaches on two real-world datasets on 1) how policy changes in Alcohol outlet licensing affect emergency service calls, and 2) how COVID19 lockdown measures affect emergency services use. Our results show that our novel method can outperform the accuracy of state-of-the-art predictions, thereby estimating the size of a causal effect more accurately. The experimental results are statistically significant, indicating our framework generates better counterfactual predictions.