Probabilistic Causal Effect Estimation With Global Neural Network Forecasting Models.

Area of Science:

• Causal inference
• Time series analysis
• Deep learning

Background:

• Estimating intervention effects on multiple units is challenging.
• Conventional methods struggle with non-mean effects and skewed distributions.

Purpose of the Study:

• Introduce DeepProbCP, a novel framework for causal effect estimation.
• Improve policy evaluation by forecasting counterfactual probability distributions.

Main Methods:

• Combine probabilistic forecasting with global deep learning (DL) models.
• Utilize autoregressive recurrent neural networks with conditional quantile functions.
• Train on stacked univariate time series for causal identification.

Main Results:

• DeepProbCP accurately forecasts counterfactual distributions.
• The framework captures effects on distribution tails and variance.
• Empirical evaluations show superior performance over state-of-the-art models.

Conclusions:

• DeepProbCP enhances causal inference by probabilistic forecasting.
• The method provides a more comprehensive understanding of intervention impacts.
• It addresses limitations of nonprobabilistic counterfactual inference methods.