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:

Using remote sensing to improve causal impact evaluation. Luke Sanford, Charles Xu, Jackson Pullman

In this paper I estimate the effects of a land tenure formalization program in the Alibori and Borgou departments of Benin on developmental and environmental outcomes. I develop a new impact evaluation approach, combining machine learning methods for satellite remote sensing data with a double machine learning approach to causal inference. The approach conditions the estimated effect on the series of satellite images taken in the two years before treatment was assigned. I validate this method by showing that I can recover estimates of effects from the randomized control trial even when I do use any information about the locations of the “control” units. I show that my approach can generate unbiased and more precise estimates at the village level, then demonstrate that it can be used to estimate the effects at the parcel level–something which has previously been impossible in land titling research without very strong assumptions. This allows me to estimate outcomes which are directly of interest to policymakers. Specifically, I find no effects of the land tenure formalization on treecover, other natural vegetation, or annual cropland. I find that there is evidence for the expansion of built up areas, but only outside of the titled parcels. I show that standard methods recover biased estimates of these effects. This method can be adapted to improve estimates the effects of many spatially targeted policies while accounting for a host of potential confounding variables common in coupled human-natural systems.

Remote Control: Debiasing Remote Sensing Predictions for Causal Inference. Matthew Gordon, Megan Ayers, Eliana Stone, Luke Sanford

Advances in machine learning and the increasing availability of satellite imagery have led to the proliferation of social science research that uses remotely sensed measures of human activity or environmental outcomes to infer the impact of policy. These measures are frequently substituted into causal inference analyses as replacements for outcome values that are not directly observable. However, if these measures suffer from prediction error that is correlated with policy variables or important confounders, as is the case for many widely used remote sensing data sets, estimates of the causal impacts of the policies may be biased. Even in a randomized controlled trial where researchers have full control over the intervention assignment mechanism, it is possible for bias to arise if the intervention or the true outcome influences the outcome prediction error. In non-experimental settings, opportunities for dependencies of this nature between the intervention variable and the outcome error multiply, related to common concerns about confounding variables in observational studies. In this project, we study how this bias arises from machine learning models that simply minimize a standard loss function, and develop an adversarial debiasing model (Zhang, Lemoine and Mitchell, 2018) in order to correct this issue when generating machine learning predictions for use in causal inference.

Sequestering Carbon through Protection and Production: A Case Study of Industrial Reforestation in Mata Atlantica, Brazil. Mark Ashton, Thomas Harris, Luke Sanford, Yuan Yao and Daniel Piotto

Many studies have examined the potential of industrial plantations to facilitate natural second growth beneath for potential conversion to native forest. Other studies have also examined the potential role of tree planting for carbon sequestration in the tropics - mostly with a focus on native trees. But the vast majority of today’s reforestation is for industrial purposes. We are not aware of any studies which have examined the potential synergies of industrial reforestation with secondary forest regrowth. Our study proposes to examine this potential by investigating: 1) the role industrial plantations can play in both elevating the carbon stored in soils and above ground biomass as compared to former and current agricultural and pastoral lands; 2) the potential for long-term carbon storage in manufactured wood products; 3) the potential for these wood products to be substitutes for more energy intensive materials; and 4) their potential to restore and protect recovering second growth within the reforestation area. Positive results from our proposed study can provide a large scalable carbon sequestration solution for reforestation globally when done correctly. This class of solutions have enormous potential because they are incentive-compatible. They feasibly improve the well-being of all actors involved–from government officials to industrial managers to landowners who lease property.

Sub-national Electoral Deforestation Cycles Luke Sanford and Eliot Carlson

In a first paper (Sanford 2021) describes the incentives that create higher rates of deforestation in election years. In a follow-up project, we extend the analysis to sub-national elections to test hypotheses about where politicians might choose to allow additional deforestation.

Discovering Persuasive Concepts Related to Renewable Energy Using Deep Neural Networks Molly Roberts, Megan Ayers, Kevin Li, Luke Sanford

Bodies of text can be powerful tools for persuading individuals, but their complex and high dimensional nature makes it difficult to distill their influential features. Neural networks are powerful tools that are able to learn latent patterns to make predictions about text data, but it is typically complicated to translate their intricate logic into meaningful insights. We develop methodology to make neural network predictions about text data more interpretable and to use their learned patterns to identify predictive text features. We use these methods to build a machine learning model which learns to predict how text influences human sentiments about government support for renewable energy, and then extract learned clusters of human-interpretable phrases which are most predictive. We utilize these phrases to experimentally test their persuasiveness in a randomized control trial. Methodologically, this paper connects research in machine learning interpretability to survey methodology. We allow researchers to discover how persuasive treatments are expressed “in the wild”, rather than relying on treatments that are subjectively specified by researchers ex ante, and provide a pipeline connecting these insightful exploratory investigations to principled confirmatory analyses.

The effects of land tenure formalization depend on the political setting–evidence from an RCT in Benin. Luke Sanford

I evaluate the impact of the Benin Plan Foncier Rural land tenure formalization program on land sparing–one of the key environmental and climate impacts that such reforms target. I find a positive and substantively important decrease in the clearing of natural vegetation through the use of fires. However, these effects are initially geographically concentrated in northern areas which have low initial levels of land tenure security, while southern and central parts of the country do not have a detectable effect. After an election in 2016 which reduced confidence in the judiciary and other formal institutions, the effect of the PFR is reduced significantly and is not distinguishable from zero in any area. This paper shows the first evidence from an RCT that political factors moderate the effects of land tenure formalization.

Determining how much carbon carbon offsets offset. (LAB)

We have early stage research in empirically estimating the amount of carbon that is truly offset when a landowner sells carbon offsets from their land. We are exploring how remote sensing should or should not be used to improve monitoring, best practices for extimating baseline scenarios, and how machine learning can help to minimize adverse selection on the part of offset sellers.

Capacity to Recover. Luke Sanford, Mark Ashton, Thomas Woodbury

In March 2019, the UN general assembly declared the years from 2021-2030 the “UN Decade on Ecosystem Restoration” to recognize the global degradation of landscapes and the need for ecosystem recovery to combat climate change and biodiversity loss. However, as we enter this decade, critical knowledge gaps remain in promoting ecosystem recovery, particularly in understudied andunder-resourced regions of the forested tropics, rich in stored carbon (Deikumah et al., 2014; Pan et al.,2011). Regenerating tropical forests in Asia have received relatively little attention from the Western-centered scientific community (Deikumah et al., 2014). The rate at which these ecosystems can recover is largely based on speculation from data collected in Neotropical forests with very different taxonomy, biogeography, land-use history, and human resource pressures (Chazdon et al., 2016; Poorter et al., 2016). In the Asian tropics, degraded forest fragments embedded within human-dominated landscapes are important for carbon sequestration (Ashton et al., 2014). They may be the most likely landscape features to return to the levels of biomass present in intact forests (Ashton et al., 2014; Osuri et al., 2014, Poorter et al. 2021). While many studies have documented degradation due to fragmentation at the global scale, the potential for recovery of forest fragments to contribute to carbon capture remains largely unexplored (Arroyo‐Rodríguez et al., 2017). As a result, current expectations for Asian tropical regenerating forests are likely inaccurate. With only 10% of remaining tropical forests housed within protected areas (Schmitt et al., 2009) and many studies reporting increased human dominance in these carbon rich ecosystems (Ashton et al., 2014; Asner et al., 2005; Lewis et al., 2015) forest fragmentation is likely to continue in the future (Hansen et al., 2020; Taubert et al., 2018). Given both their continued generation and the potential of forest fragments for restoration, understanding the long-term dynamics of forest recovery and carbon capture is vital for informing future management.

This study will assess the dynamics of Asian tropical forest fragments surrounding the Sinharaja Man and Biosphere Reserve in south western Sri Lanka. An area largely representative of the mixture of Dipterocarp forest fragments,agriculture, and human settlement prevalent across human-dominated landscapes in tropical wet Asia. The landscape surrounding Sinharaja is a mixture of tea plantations, tree plantations, home gardens, rice paddies, degraded forests, and scrublands (Martin et al., 2019) that have shifted over time based on agricultural markets and social values (Caron, 1995). Recent legislation by the outgoing president will quadruple the reserve’s size by incorporating surrounding forests (Rodrigo, 2019). However, selective logging and small-holder conversion for agriculture within these forests has led to a patchwork of forest fragments of different ages. Previous extensions of the forest reserve also occurred in both 1992 and 2003. This has created a unique setting which allows us to study the long-term dynamics of forest fragments and how landscape dynamics and recovery are affected by non-forest land uses in the mosaic landscape.

Evaluating the causes and consequences of Kenyan resettlement schemes using satellite imagery and deep learning. Richard Bluhm, Dannie Daley, Luke Sanford

Indian Legal Responsiveness to Environmental Issues–Data from 5 Million Court Cases. Luke Sanford and Teevrat Garg

Predicting Amazonian Deforestation with Political, Economic, Social, Natural, and Remote Sensing data. With Annie Ulichney and Kathryn Baragwanath