One number, fully loaded

“For the world as a whole, its top 5% percent of polluters contributed 73% of all electricity-based CO₂ discharges or 14.6 times more than if pollutants were evenly dispersed.”

That sentence is quoted, not paraphrased. Anyone can click through and check me — that is the point of this page.

The paper is by Don Grant, David Zelinka, and Stefania Mitova (Sociology, University of Colorado Boulder), funded by the National Science Foundation (award #1824106). It went through peer review at Environmental Research Letters: received 7 June 2021, revised 7 July, accepted 13 July, published 24 August 2021.

My part was the climate-data engineering underneath it — the case study walks through that work: integrating 16 databases, geocoding 30,000+ plants, and gap-filling what remained.

The number rests on an updated edition of CARMA (Carbon Monitoring for Action), the most widely used bottom-up inventory of power-plant CO₂ emissions. The 2018 edition draws on plant-level emissions reports from official registries — the U.S. (eGRID), the European Union (the European Pollutant Emission Register), Australia, Canada, and India — plus global plant- and company-level data from Platts.

Getting from those scattered sources to one analyzable table is most of the work: facility records arrive with inconsistent identifiers, missing coordinates, and gaps in the emissions and generation figures. Reconciling that is what the ETL, geocoding, and spatial-join pipeline in the case study did.

Not every plant in the world reports its emissions. Where reports were missing, values were estimated with gradient-boosting models (Friedman's GBM, via the H2O.ai platform), fit with ten-fold cross-validation on the plants that do report: 3,019 plant-level observations with an observable capacity factor and 2,581 with an observable CO₂ emission factor.

An honest number tells you where it is soft, so: model-filled values are estimates, not measurements. The paper itself flags that where predicted emissions run close together, the relative ranking of individual top plants is less certain. The 73% is a distributional finding across tens of thousands of plants — it does not depend on any one plant's exact rank.

A number this concentrated changes the policy question. The paper estimates that 17%–49% of the world's electricity-generation CO₂ emissions could be eliminated, depending on whether the hyper-emitting plants were held to intensity standards, switched fuels, or added carbon capture — targeted instruments, rather than economy-wide ones alone.

And that is what provenance buys: when someone challenges the 73% — a regulator, a reviewer, an auditor, opposing counsel — every layer above answers them. This is the standard everything I deliver is held to. If your numbers can't answer these questions yet, that's the problem I work on.