Recent Working Papers
Distributed Effects of Climate Policy: A Machine Learning Approach
CEEPR Working Paper WP-2020-R3
Tomas Green (Energy Futures Initiative) and Christopher R. Knittel (MIT)
We employ machine learning techniques to estimate household carbon footprints (HCFs) for the average household in each Census tract—geographic areas that represent roughly 4,000 people. We find that there is significant variation in carbon footprints across income and geography; income effects are driven by higher footprints related to transportation and consumer products and services, while geographic effects are primarily a result of the variable carbon intensity of the electricity grid. Using these footprints, we assess the net effects of various climate policies on households in the United States paying particular attention to the distribution across geography, urbanity, and income groups. Our objective is to improve the understanding of the potential for regressivity, geographic transfers, and rural-urban transfers among climate policy options and test for ways to control for transfers—preserving transfers from high-income households to low-income households, but mitigating transfers from rural areas to urban areas and from the Midwest and South to the Coasts. Our focus is on the net increase or decrease of annual household expenses under 12 different policy scenarios, which included both carbon pricing schemes and regulatory standards. We find regulatory standards tend to be regressive and, on average, are a net cost to low-income households—especially those in rural areas. Carbon pricing, when accompanied with a dividend, is progressive for urban, rural, and suburban households, with the average low-income household receiving a larger dividend check than they spend in carbon taxes. However, there are transfers from the Midwest and Plains to the Coasts when the dividend is evenly divided. We show that this can be mitigated through adjusting the dividend slightly (<8% increase or decrease). Increasing the progressive structure of a policy benefits rural households more on average, but increases the overall heterogeneity of impacts within each income group. Reducing the transfers between geographic regions and urban-rural households increases the average benefit to low-income households and reduces the heterogeneity of impacts within income groups. We encourage policy makers to assess and control for unwanted transfers between households
Driving Behavior and the Price of Gasoline: Evidence from Fueling-Level Micro Data
Journal of Public Economics, October 2021
Christopher R. Knittel (MIT) and Shinsuke Tanaka (Tufts)
We use novel microdata on on-road fuel consumption and prices paid for fuel in Japan to estimate short-run price elasticities of demand for gasoline consumption. We have three main findings. First, our elasticity estimates of roughly -0.37 are in orders of magnitude larger than previously estimated using more aggregate data. Second, we are one of the first papers to separately estimate both the price elasticities of kilometers driven (-0.30) and on-road fuel economy (0.07). Lastly, we find that on-road fuel economy is determined by recent prices than distant past prices paid, suggesting limited habit formation of fuel-conserving driving behaviors.
Using Machine Learning to Target Treatment: The Case of Household Energy Use
Revisions requested from The Economic Journal
Christopher R. Knittel (MIT) and Samuel Stolper (Michigan)
We use causal forests Athey et al (2018) to evaluate heterogenous treatment effects within a series of large-scale randomized experiments. Our application is the Home Energy Report, a widely-used behavioral nudge that relies on repeated social comparison to encourage household energy efficiency. Our data consist of monthly electricity consumption, treatment assignment, and cross-sectional characteristics of 902,581 New England households. The program-wide average treatment effect is a monthly reduction in electricity consumption of approximately 9 kilowatt-hours (kWh), or 1 percent; the full range of type-specific impacts, however, runs from -30 to +10 kWh. The forest algorithm uses two variables with disproportionately high frequency in developing predictions: pre-treatment consumption and home value. In addition, we find evidence of a ``boomerang effect'': households with relatively lower consumption than neighbors of the same type are the ones for which we estimate positive treatment effects. Finally, using forest results to selectively target treatment increases annual social welfare by 12-120 percent, depending on the program year and parameterization of the welfare function.
The Use of Regression Statistics to Analyze Imperfect Pricing Policies
The Journal of Political Economy, May 2020
Mark R. Jacobsen (UCSD), Christopher R. Knittel (MIT), James M. Sallee (Berkeley), Arthur A. van Benthem (Wharton)
Corrective taxes can completely solve a variety of market failures, but actual policies are commonly forced to deviate from the theoretical ideal due to administrative or political constraints. This paper presents a method that requires a minimum of market information to quantify the efficiency costs of constraints on the design of externality-correcting tax schemes, or more generally the costs of imperfect pricing, using simple regression statistics. We demonstrate that, under certain intuitive conditions, standard output from a regression of true externalities on policy variables, including the R2 and the sum of squared residuals, has an immediate welfare interpretation—it characterizes the relative welfare gains achieved by alternative policies. We utilize our approach in four diverse empirical applications: random mismeasurement in externalities, imperfect spatial policy differentiation, imperfect electricity pricing, and heterogeneity in the longevity of energy-consuming durable goods. In two cases, we find that policy constraints are relatively harmless, while in the other two cases, the constraint induces large inefficiencies. Regarding the case of durable longevity, we find that policies that regulate vehicle fuel economy, but ignore the differences in average longevity across types of automobiles, recover only about one-quarter to one-third of the welfare gains achievable by a policy that also takes product longevity into account.
Attribute Substitution in Household Vehicle Portfolios
The Rand Journal of Economics, December 2020
James Archsmith (UC Davis), Kenneth Gillingham (Yale), Christopher R. Knittel (MIT), and David Rapson (UC Davis)
Household preferences for goods with a bundle of attributes may have complex substitution patterns when one attribute is changed. For example, a household faced with an exogenous increase in the size of one television may choose to decrease the size of other televisions within the home. This paper quantifies the extent of attribute substitution in the context of multi-vehicle households. We deploy a novel identification strategy to examine how an exogenous change in the fuel economy of a kept vehicle affects a household's choice of a second vehicle. We find strong evidence of attribute substitution in the household vehicle portfolio. This effect operates through car attributes that are correlated with fuel economy, including vehicle footprint and weight. Our findings suggest that attribute substitution exerts a strong force that may erode a substantial portion of the expected future gasoline savings from fuel economy standards, particularly those that are attribute-based. Elements of our identification strategy are relevant to a broad class of settings in which consumers make sequential purchases of durable portfolio goods.