GMEF v1.0 - Global Marginal Emission Factor panel for electricity, 1999–2025

An open panel of marginal emission factors for electricity - the greenhouse-gas intensity of an additional kilowatt-hour of demand - covering 203 economies from 1999 to 2025.

Decisions that add electricity load (electric vehicles, heat pumps, electrolysers, data centres) change emissions by the intensity of the generation that responds, not by the average intensity of the system. Average grid factors are nevertheless the only figures published for most countries. Empirically estimated marginal factors exist for the United States, for parts of Europe, and as proprietary or projection-based products covering a few dozen countries; for most of the world they do not exist. GMEF fills that gap from fully public inputs.

For every economy-year the dataset reports an average emission factor; UNFCCC Tool 07-style operating, build and combined margins; and three econometric marginal factors with bootstrap confidence intervals — together with a technology-level decomposition of which sources actually serve incremental demand. The preferred estimator treats the margin as a share vector rather than a single slope: because demand equals generation plus net imports as an accounting identity, marginal technology shares sum to one by construction, and projecting them onto the probability simplex yields a factor bounded inside the convex hull of the technology factors combined, decomposable by technology, and stable in small systems.

Validation: the structural decomposition reproduces the reduced-form slope at r = 0.9997; annual and monthly estimators built from different releases and different differencing schemes agree at r = 0.95 among high-quality economies; EPA AVERT (744 gCO₂/kWh) and eGRID non-baseload (703) both fall inside the estimated US interval [672, 950]; and India's Central Electricity Authority build margin for FY2023-24 (552) is matched to within 9 gCO₂e/kWh.

Built entirely from Ember's Yearly and Monthly Electricity Data releases (CC-BY-4.0) and fully regenerable from public sources by rerunning the seven included scripts. Bootstrap seed 20260906; results are exactly reproducible.