Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Spatiotemporal double machine learning to estimate the impact of Cambodian land concessions on deforestation

Anika Arifin, Duncan DeProfio, Layla Lammers, Benjamin Shapiro, Brian J Reich, Henry Uddyback, Joshua M Gray

TL;DR

This work conducts a large-scale simulation study using Bayesian Additive Regression Trees with spatial embeddings and finds that, under certain conditions, the DSR model outperforms standard approaches for addressing unobserved spatial confounding.

Abstract

Environmental policy evaluation frequently requires thoughtful consideration of space and time in causal inference. We use novel statistical methods to analyze the causal effect of land concessions on deforestation rates in Cambodia. Standard approaches, such as difference-in-differences regression, effectively address spatiotemporally-correlated treatments under some conditions, but they are limited in their ability to account for unobserved confounders affecting both treatment and outcome. Double Spatial Regression (DSR) is an approach that uses double machine learning to address these scenarios. DSR resolves the confounding variables for both treatment and outcome, comparing the residuals to estimate treatment effectiveness. We improve upon DSR by considering time in our analysis of policy interventions with spatial effects. We conduct a large-scale simulation study using Bayesian Additive Regression Trees (BART) with spatial embeddings and find that, under certain conditions, our DSR model outperforms standard approaches for addressing unobserved spatial confounding. We then apply our method to evaluate the policy impacts of land concessions on deforestation in Cambodia.

Spatiotemporal double machine learning to estimate the impact of Cambodian land concessions on deforestation

TL;DR

This work conducts a large-scale simulation study using Bayesian Additive Regression Trees with spatial embeddings and finds that, under certain conditions, the DSR model outperforms standard approaches for addressing unobserved spatial confounding.

Abstract

Environmental policy evaluation frequently requires thoughtful consideration of space and time in causal inference. We use novel statistical methods to analyze the causal effect of land concessions on deforestation rates in Cambodia. Standard approaches, such as difference-in-differences regression, effectively address spatiotemporally-correlated treatments under some conditions, but they are limited in their ability to account for unobserved confounders affecting both treatment and outcome. Double Spatial Regression (DSR) is an approach that uses double machine learning to address these scenarios. DSR resolves the confounding variables for both treatment and outcome, comparing the residuals to estimate treatment effectiveness. We improve upon DSR by considering time in our analysis of policy interventions with spatial effects. We conduct a large-scale simulation study using Bayesian Additive Regression Trees (BART) with spatial embeddings and find that, under certain conditions, our DSR model outperforms standard approaches for addressing unobserved spatial confounding. We then apply our method to evaluate the policy impacts of land concessions on deforestation in Cambodia.
Paper Structure (14 sections, 16 equations, 3 figures, 6 tables, 1 algorithm)

This paper contains 14 sections, 16 equations, 3 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Cambodian deforestation background and data
  3. Study Area
  4. Deforestation, Concession and Ancillary Data
  5. Double machine learning for spatial difference-in-differences
  6. Stage 1: Spatial regressions
  7. Stage 2: Final estimate and standard error
  8. Extension to clustered treatments
  9. Simulation study
  10. Simulation design with pixel-level treatment allocation
  11. Simulation design with clustered treatment allocation
  12. Results
  13. Application to Cambodian deforestation
  14. Discussion

Figures (3)

  • Figure 1: Deforestation and concessions: Deforestation by pixel in the pre-intervention (left) and post-intervention (right) period. Purple represents little deforestation, while brighter greens show significant deforestation. A yellow result represents 100 percent deforestation in the raster. A concessions graph is in the center, where yellow indicates the conceded land.
  • Figure 2: Treatment effect estimation using varying levels of spatial information, from no spatial information to $L$ = 196 Wendland basis functions. The points are estimates, and bars are 95$\%$ confidence intervals.
  • Figure 3: Simulation study comparing OLS, DID, three methods without random effects, and three methods with random effects. The true value of $\gamma=3$, and smoothness $\nu=2$.