Prices and efficiency policy stand out in a study of U.S. energy use

Energy use depends on more than how much an economy produces or how cold its winters are. Prices, infrastructure, technology and demand for comfort also matter. A preprint by David Benatia, Rémy Molinié and Pierre-Olivier Pineau examines how these factors relate to energy consumption across the 50 U.S. states and the District of Columbia.

The study covers 2006–2022, with 867 state–year observations. Its main energy measure includes electricity-system losses allocated to end users, rather than counting only energy delivered to homes, businesses and vehicles. The findings are preprint results, not established causal estimates.

A national decline hides different state stories

The researchers first use an accounting method called Logarithmic Mean Divisia Index decomposition to separate changes in per-person energy use into activity, energy intensity and weather-related contributions.

National consumption per person fell 12.8% over the period. Lower energy intensity—energy used relative to indicators such as floor area, industrial output or vehicle travel—more than accounts for that decline. Growing residential and commercial floor space partly offset the reductions, while the combined climate contribution was small.

This accounting does not establish why intensity improved. Its intensity category captures changes left after the specified activity and climate factors are accounted for; it is not a direct measurement of technological efficiency.

State experiences also differed. North Dakota combined a large expansion in industrial activity with relatively modest intensity declines. Louisiana experienced falling activity offset by rising industrial intensity. Similar overall consumption trends can therefore conceal different underlying changes.

Separating a benchmark from a gap

To examine differences in consumption levels, the authors estimate a stochastic-frontier model. Its “frontier” is a conditional economic benchmark: the energy needed to meet reference service needs using a lifetime-cost-minimizing technology mix, given prices and state characteristics. It is not a physical minimum dictated by engineering.

Observed consumption can exceed that benchmark through a composite gap. This can include older technology, barriers to investment, additional comfort-driven consumption and rebound, where improved efficiency encourages greater use of energy services. A separate noise component allows for measurement error and model approximation.

The distinction rests on assumptions. The baseline statistical model assumes a half-normal distribution for the nonnegative gap. The theoretical framework also requires an aggregate condition ensuring that technology-related and above-baseline consumption outweigh energy reductions associated with unmet service needs. The gap is therefore neither a pure measure of waste nor a measure of welfare loss.

What receives the largest statistical shares?

The authors adapt a relative-importance method that averages predictors’ contributions across all possible orders of entry into an auxiliary regression. They combine separate allocations for the fitted frontier and generated gap.

Under their accounting convention, 56.3% of variation in logged per-person energy use is assigned to the frontier, 42.2% to the composite gap and 1.5% to noise. These are model-dependent variance shares, not percentages of national energy consumption. The calculation divides a covariance residual equally between the frontier and gap; that division is not unique.

Efficiency policy receives a combined allocation of about 15.8%, narrowly ahead of prices at 15.1%. Together they account for roughly 31% of total log-energy variation. Prices contribute mainly through the frontier; policy contributes through both channels. Vehicle travel and fossil-fuel production are the next leading observed contributors.

Policy and price rankings remain broadly prominent across 21 specifications, but neither occupies a fixed rank in every model. Their sampling-uncertainty intervals also overlap substantially. Most generated-gap variation remains unexplained by the three gap predictors: policy, prices and GDP per person.

Association is not a policy experiment

The baseline estimates combine persistent differences between states with changes within them. Separating those sources produces important contrasts. The frontier price coefficient is about −0.78 between states but only −0.17 within states.

For policy, the frontier association reverses sign: states with persistently higher scores have lower fitted benchmarks, while increases relative to a state’s own average accompany higher benchmarks. Policy responses to energy pressures and delayed effects are possible explanations, not mechanisms the study identifies.

Measurement matters too. The policy score combines programs, standards, institutions and selected performance indicators; it does not directly measure realized savings, and its rubric changes over time. Some findings also change with the energy outcome or assumed gap distribution.

The authors recommend durable price signals, coherent efficiency programs, structural investment and targeted household support. Their analysis supports investigating these margins together, but does not establish how much energy any particular intervention would save.