UC Berkeley

As demand for electricity grows around the world, so does the need for rigorous evaluation of energy policy interventions. In this dissertation, I use large datasets and modern econometric methods to study two such policies at scale - rural electrification in India and energy efficiency subsidies in California. I find that the benefits associated with these interventions are substantially smaller than previously thought, highlighting the importance of using new techniques for causal inference in these settings. In Chapter 1, I study the impacts of grid-scale rural electrification in India, using a regression discontinuity framework. In Chapter 2, I evaluate energy efficiency upgrades in K-12 schools in California using high-frequency data and novel machine learning methods. In Chapter 3, I develop methods to guide experimental design in the presence of panel data.

In the first chapter, coauthored with Louis Preonas, we study the impacts of energy access in the developing world. Over 1 billion people still lack electricity access. Developing countries are investing billions of dollars in rural electrification, targeting economic growth and poverty reduction, despite limited empirical evidence. We estimate the effects of rural electrification on economic development in the context of India's national electrification program, which reached over 400,000 villages. We use a regression discontinuity design and high-resolution geospatial data to identify medium-run economic impacts of electrification. We find a substantial increase in electricity use, but reject effects larger than 0.26 standard deviations across numerous measures of economic development, suggesting that rural electrification may be less beneficial than previously thought.

In the second chapter, coauthored with Christopher R. Knittel, David Rapson, Mar Reguant, and Catherine Wolfram, we study the impacts of energy efficiency investments at public K-12 schools in California. We leverage high frequency data -- electricity use every 15 minutes -- to develop several approaches to estimating counterfactual energy consumption in the absence of the efficiency investments. In particular, We use difference-in-differences approaches with rich sets of fixed effects. We show, however, that these estimates are sensitive to the set of fixed effects included and to the set of schools included as controls. To address these concerns, We develop and implement a novel machine learning approach to predict counterfactual energy consumption at treated schools and validate the approach with non-treated schools. We find that the energy efficiency projects in our sample reduce electricity consumption between 2 to 5% on average, which can result in substantial savings to schools. We also compare our estimates of the energy savings to ex ante engineering estimates. Realized savings are generally less than 50% of ex ante forecasts and quite low for measures other than heating and air-conditioning systems or lighting.

In the third chapter, coauthored with Louis Preonas and Matt Woerman, we seek to answer: How should researchers design experiments with panel data? We derive analytical expressions for the variance of panel estimators under non-i.i.d. error structures, which inform power calculations in panel data settings. Using Monte Carlo simulation, data from a randomized experiment in China, and high-frequency U.S. electricity consumption data, we demonstrate that traditional methods produce experiments that are incorrectly powered with proper inference. Failing to account for serial correlation yields overpowered experiments in short panels and underpowered experiments in long panels. Our theoretical results enable us to achieve correctly powered experiments in both simulated and real data.