Heterogeneous Treatment Effects and Counterfactual Policy Targeting Using Deep Neural Networks: An Application to Central Bank Corporate Credit Facilities

Abstract

I present a novel two-step semi-parametric difference-in-differences (DiD) estimator for computing dynamic (heterogeneous) treatment effects and policy counterfactuals. In the first step, deep neural networks are used to compute non-parametric terms in a setting with high-dimensional controls. These are inputs into the estimator evaluated in the second step. The estimator is applied to study the effects of the Federal Reserve's Corporate Credit Facilities (CCFs) on the dynamics of firm cash holdings, leverage, payout, and investment. I show that the proposed estimator delivers comparable results to static (homogeneous) treatment effects obtained from DiD panel regressions and dynamic (homogeneous) treatment effects from event study regressions with two-way fixed effects, though with important differences attributable to selection bias and heterogeneity. Firms generally increased cash holdings and leverage, while payout and investment initially fell. Firms eligible for the CCFs accumulated less cash and began deleveraging in 2021, relative to ineligible firms. Eligible firms exhibit relatively larger payouts, while they do not invest more, suggesting that the CCFs failed to meet their objective of boosting real effects. Counterfactual eligibility criteria can possibly deliver an improvement in investment outcomes.

Keywords: fed ccfs, neural networks, macro-finance