Estimating Continuous Treatment Effects in Panel Data using Machine Learning with a Climate Application

📅 2022-07-18
📈 Citations: 1
Influential: 0
📄 PDF

career value

196K/year
🤖 AI Summary
This paper addresses model misspecification in two-way fixed effects (TWFE) estimators under nonlinear continuous treatment effects. We propose a semiparametric estimator for panel data that enables unbiased identification of the average partial derivative (APD). Methodologically, we extend the double/debiased machine learning (DML) framework—previously developed for cross-sectional settings—to continuous-treatment panel models with unit fixed effects, integrating high-dimensional regressions (e.g., Lasso or random forests) and robust standard error construction, and formally establish the asymptotic normality of the estimator. Empirically, we apply the method to estimate the impact of extreme heat on maize yields. Results reveal a nonlinear, marginally diminishing dose–response relationship: linear TWFE estimates underestimate heat-induced yield losses by 50%, implying an additional annual loss of USD 3.17 billion by 2050.
📝 Abstract
This paper introduces and proves asymptotic normality for a new semi-parametric estimator of continuous treatment effects in panel data. Specifically, we estimate the average derivative. Our estimator uses the panel structure of data to account for unobservable time-invariant heterogeneity and machine learning (ML) methods to preserve statistical power while modeling high-dimensional relationships. We construct our estimator using tools from double de-biased machine learning (DML) literature. Monte Carlo simulations in a nonlinear panel setting show that our method estimates the average derivative with low bias and variance relative to other approaches. Lastly, we use our estimator to measure the impact of extreme heat on United States (U.S.) corn production, after flexibly controlling for precipitation and other weather features. Our approach yields extreme heat effect estimates that are 50% larger than estimates using linear regression. This difference in estimates corresponds to an additional $3.17 billion in annual damages by 2050 under median climate scenarios. We also estimate a dose-response curve, which shows that damages from extreme heat decline somewhat in counties with more extreme heat exposure.
Problem

Research questions and friction points this paper is trying to address.

Estimating nonlinear continuous treatment effects in panel data
Addressing bias in average partial derivative estimation
Developing machine learning methods for climate impact analysis
Innovation

Methods, ideas, or system contributions that make the work stand out.

Automatic double machine learning for panel data
Optimization debiasing with analytic derivatives
Handles nonlinear relationships and fixed effects
🔎 Similar Papers