Website CV
Fields
Environmental & Energy, Agricultural, Urban & Regional
Job market paper
Identifying Counterfactual Solar PV Sites with CNN and Propensity Score Matching: Effects on Land Transactions
Utility-scale solar photovoltaic (PV) capacity in the U.S. Midwest grew from 10 MW in 2009 to over 23 GW in 2024, converting more than 86,000 acres, mostly former cropland and pasture. Existing estimates of solar’s effects on nearby land values conflict: Abashidze and Taylor (2023) find no spillover to farmland prices, while Hu et al. (2025) report a 19% premium within two miles. This inconsistency may reflect site-selection endogeneity: both treated parcels and distance-based controls are defined relative to locations developers selected. I train a convolutional neural network on a 25-channel geospatial image stack of siting-relevant features to predict solar-suitable but undeveloped locations, estimate development propensity scores using grid and electricity-market data, and match each realized Midwest site to similar undeveloped sites. I link realized and counterfactual sites to land transactions within 5 km, cleaned with standard transaction filters from prior studies, and estimate stacked difference-in-differences models over three years before and after construction. Pre-treatment trends are flat, supporting the parallel-trends assumption. Preliminary results show no significant average effect of PV construction on nearby farmland or large-lot residential prices across specifications; large-lot prices per acre rise transiently in the first two post-construction years before fading. A triple difference by distance indicates that parcels within 3 km of a built site sell at a modest discount relative to parcels 3-5 km away, the opposite of the reported near-site premium. These findings suggest positive estimates in prior work partly reflect siting selection, with implications for siting policy and rural compensation design.