Broadly, we work on how political institutions shape incentives to preserve the environment or exploit natural resources. We do this using new, high-dimensional sources of data including satellite imagery and human-generated text. This data allows us to generate insights at a local scale for large geographic areas or populations. Some of our ongoing projects are described below:
Adversarial Debiasing for Unbiased Parameter Recovery
Luke C Sanford, Megan Ayers, Matthew Gordon, Eliana Stone
Under Review, February 2025.
Advances in machine learning and the increasing availability of high-dimensional data have led to the proliferation of social science research that uses the predictions of machine learning models as proxies for measures of human activity or environmental outcomes. However, prediction errors from machine learning models can lead to bias in the estimates of regression coefficients. In this paper, we show how this bias can arise, propose a test for detecting bias, and demonstrate the use of an adversarial machine learning algorithm in order to de-bias predictions. These methods are applicable to any setting where machine-learned predictions are the dependent variable in a regression. We conduct simulations and empirical exercises using ground truth and satellite data on forest cover in Africa.
arXiv:2502.12323
Elevating Spatial Evaluation: Satellite-Driven Confounder Adjustment
Luke Sanford
Revise and Resubmit, 2025.
Estimating causal effects of geospatial interventions, such as the presence of certain policies, on environmental, development, and conflict outcomes poses significant challenges. Traditional causal inference methods often fail to account for complex, hard-to-measure confounders in these settings. This paper proposes using satellite imagery and machine learning to adjust for confounders using a double machine learning approach. I demonstrate its effectiveness by conducting a Lalonde-style replication of findings from a land tenure RCT in Benin. The satellite double machine learning (S-DML) method recovers RCT treatment effects across multiple outcomes where traditional observational approaches fail. This approach offers several advantages: it leverages free satellite data, provides a computationally efficient way to process imagery, and enables evaluation of historical interventions. The S-DML method has broad potential applications in evaluating environmental policy, conflict, economic growth, and other spatial treatments where confounders may appear in satellite imagery.
SSRN Paper
Causal Carbon: Baselines and Additionality with Potential Outcomes
Megan Ayers, Luke Sanford, Will Gardner, Sara Kuebbing
OSF Preprints, February 2025.
Recent work has questioned the credibility of forest carbon offsets as an environmental intervention and nature-based solution for mitigating climate change. Despite some updates to carbon credit methodologies and advice to purchase only high-integrity or high-quality credits, it is not clear which carbon offsets meet these standards under which conditions. In this paper, we draw on the fields of statistics and causal inference to develop a generalized framework for analyzing carbon offset protocols. We show that strategic enrollment combined with even seemingly innocuous measurement errors in carbon stocks can lead to market distortions and that there is an inherent tradeoff between minimizing these distortions and broadening enrollment. The provided framework clarifies what purchasers of carbon offsets must believe about the world in order for purchased credits under each protocol to accurately reflect the impact of crediting programs and builds common ground on which more fruitful engagement between different sectors of the carbon market can build agreement.
Google Scholar Citation
Strategic behavior in jurisdictional REDD+
Alberto Garcia, Luke Sanford
SSRN Working Paper, November 2024.
While carbon markets demonstrate potential to direct climate finance to low-cost, high-impact mitigation opportunities, recent research has cast doubt on their ability to contribute to real emissions reductions. As a result, the market has shifted toward a jurisdictional approach to the generation of forest carbon credits. Here we explore whether there is potential for or evidence of behavior that undermines the integrity of the carbon credits generated in jurisdictional REDD+. We find that jurisdictional governments could make strategic enrollment decisions to maximize their generated credits (and damage the “additionality” of those credits), there is little evidence that they have done so to date. Under existing jurisdictional protocols governments that are already making progress on reducing deforestation are most incentivized to enroll, while those in most need of support face the most difficulty in attaining climate finance. Furthermore, we find that in approximately one-third of enrolled jurisdictions, deforestation temporarily increases near the start of the crediting period. We suspect this may be due to an anticipatory response on the part of landowners within soon-to-be enrolled jurisdictions. This paper shows that Jurisdictional REDD+ protocols face some challenges similar in nature to project-based REDD+ and some new challenges specific to the jurisdictional setting.
Google Scholar Citation
Large but Restricted Tree Cover Loss Associated with Mexico’s Mayan Train from 2020 to 2024
Luke Sanford, Sarah E. Castle, Cesar B. Martinez-Alvarez
Working Paper, 2025.
What are the environmental consequences of large-scale infrastructure projects in the tropics? While agriculture and climate change are well-recognized drivers of deforestation, we know less about how transportation infrastructure—particularly railways—affects, directly and indirectly, the patterns of land use change in the tropics. In this paper, we analyze the case of the Tren Maya (Maya Train) railway megaproject in Mexico’s Yucatán Peninsula using an event study research design and high-resolution remote sensing data. We employ the Continuous Change Detection and Classification (CCDC) algorithm and a difference-in-differences design to estimate quarterly deforestation before and after construction for the constructed and planned, but unconstructed, railway routes. We find that exposure to the construction of Tren Maya is associated with a large but short-lived spike in forest cover loss, particularly for areas very close to the railroad and in proximity to the train’s stations. However, we also show that, in contrast to other forms of transportation infrastructure, for example highways, the environmental damage associated with the train remained concentrated in a narrow corridor, with minimal spillover effects. Our results are robust to employing different specifications, including alternative specifications of control units and estimators. Our findings suggest that infrastructure is a crucial driver of forest loss in tropical, mega-diverse countries and that it is necessary to understand the similarities and differences among different types of transportation projects. Moreover, we show how real-time satellite monitoring and spatial causal inference can improve the oversight of mega-projects in the tropics.
The Political Economy of Conflict and Deforestation: Evidence from the Philippines
Dotan Haim, Nina McMurry, Luke Sanford
Working Paper, 2025.
Tropical forests are disappearing at alarming rates, especially in areas where armed groups challenge state authority. Nearly 40 percent of the world’s tropical forests are located in regions affected by violent conflict, yet the role of non-state armed actors in shaping forest loss remains poorly understood. We argue that rebels face a strategic tradeoff between extracting profit and preserving local legitimacy. On one hand, they can benefit from protecting illegal logging operations and taxing firms that operate in forested areas. But the same marginalized populations they rely on for support are often the most harmed by deforestation, either because it violates cultural beliefs or leads to environmental degradation that threatens safety and economic security. Using annual village-level military intelligence reports on rebel presence in the Philippines and satellite imagery to measure deforestation, we find support for this view. Rebel presence increases forest loss overall, but this effect is limited to areas without Indigenous populations and where environmental risks like landslides are low. These results highlight how armed group behavior responds to the preferences and vulnerabilities of local civilians.
Electoral Deforestation Cycles: A Subnational Analysis
Eliot L. Carlson, Luke Sanford
Working Paper, 2025.
The phenomenon of electoral deforestation cycles - in which accelerations in deforestation rates are observed during and immediately preceding competitive election years - is both well documented qualitatively and, more recently, corroborated quantitatively with remote sensing methods using satellite imagery. However, existing work either consists of single-country studies that provide evidence that politicians spatially target deforestation to competitive districts or cross-national studies that compare national-level rates of deforestation. While often compared, these groups of papers test fundamentally different theories about electoral deforestation. Are close national elections the main driver of electorally motivated deforestation? Does electorally motivated deforestation occur in competitive districts even when national elections are not close? What are the political conditions that facilitate this driver of deforestation? We combine satellite-derived global deforestation data with geographically referenced data from lower-house legislative elections. To improve global coverage, we extend the number of elections present in the geo-referenced electoral database by a factor of 10 by coding the years in which redistricting occurred in every country for which at least one geo-referenced dataset was available. We investigate how electoral competitiveness drives deforestation at both district and national levels. We find that both competitive district elections and competitive national elections are associated with deforestation, with the highest rates of deforestation occurring in competitive districts and competitive national elections. However, in non-clientelist democracies, we find that in non-competitive national elections, deforestation is lower in non-competitive districts and higher in competitive districts. This demonstrates that previous cross-national work on electoral deforestation cycles likely under-estimated the magnitude of electoral deforestation cycles by ignoring district-level returns and that these cycles exist in non-clientelist electoral democracies, albeit in a different form.
The Political Contours of Land Reform: Conditional Impacts of Tenure Formalization on Socio-Economic and Environmental Outcomes
Luke Sanford
Under Review, July 2025.
When the state introduces new formal institutions, how does people’s behavior change? The impact of institutional change depends on peoples’ relationships with informal institutions and the state. I explore this argument in the context of land tenure formalization in Benin, exploiting a large randomized control trial. I combine high-resolution remote sensing data on land use with data on informal institutions, conflict, political views, ethnicity, and geography to estimate the program’s effects on land sparing (avoided deforestation and increased agricultural productivity) across political contexts. The formalization reduced land sparing in areas with strong existing informal institutions or low trust in courts and increased land sparing in areas with high levels of land conflict and high confidence in courts. This paper uses satellite data and a large RCT to highlight how the effectiveness of state-led institutional change depends on the local institutional context and the political situation of subjects.
For the most up-to-date list of recent publications and working papers, please visit Luke Sanford’s Google Scholar profile.