Association for Public Policy Analysis & Management

In recent years, state governments have played large roles in promoting residential solar installations through a myriad of policies, which are important for mitigating climate change and transitioning to sustainable energy systems. This paper evaluates the heterogeneous effectiveness of state-level solar rebate programs at promoting residential solar energy systems in the U.S. In particular, I examine how the impacts of solar rebate programs differ by the presence of other solar policies and demographic characteristics using a machine learning method.

This paper contributes to the growing body of literature examining the determinants of increased residential solar photovoltaic (PV) capacity in the U.S. Previous studies mostly focus on the effectiveness of one specific policy, such as Solar Renewable Energy Credit and Renewable Portfolio Standards. Others have studied non-policy factors that affect residential solar adoption, such as demographic characteristics and peer effect. However, prior analyses mostly overlook the interactions among policies and non-policy factors, which can be important for effective policy implementation. For instance, the combination of well-designed interconnection and net metering standards can provide good support for distributed generation PV systems. We often do not have a complete understanding of these interrelations. Hence, empirically testing a model with a set of predetermined interaction terms might omit some important relations and yield misleading results.

This paper examines the heterogeneous effects of solar rebate programs conditioning on the implementation of other solar policies, such as interconnection and net metering standards, as well as demographic factors. I construct a zip code level panel dataset from 2005 to 2015, and control for demographic characteristics, solar resources and political climate towards pro-environmental policies. I use year fixed effect to account for the unobserved factors that are common across regions in a given year and changes in federal level policies. I utilize casual regression trees, a data-driven approach, to partition data into subgroups that differ in the magnitude of program impact. This machine learning tool can fully explore the potential interactions and uncover the variables that influence the effectiveness of rebate programs.

The innovation of this paper lies in the application of machine learning in evaluating residential solar policies in the U.S. First, I evaluate the heterogeneous impacts of solar rebate programs based on a panel dataset of 50 states, whereas previous studies mostly focus on California or a region in the U.S. Second, this is one of a few studies that apply machine learning tools to solar policy evaluation. The results provide insights on the necessary supporting policies for rebate programs and the sub-populations in which the rebates are most effective.