Census Bureau publishes local air‑conditioning estimates using cross‑survey modeling

Jay Sawyer, technical lead for model data product development at the U.S. Census Bureau, said the agency published the Local Air Conditioning Estimates, or LACE, last week and that the dataset shows the share of people with and without air conditioning at national, state, county and census‑tract levels.

The LACE tables were created by applying a cross‑survey modeling workflow that trains a machine‑learning model on a detailed subject matter survey — in this case, the American Housing Survey (AHS) — and then applies predictions to the broader American Community Survey (ACS). "We would utilize the data in the detailed subject matter survey to go ahead and train a machine learning algorithm that could be applied to the survey that had a more robust geographic sample," Sawyer said.

Sawyer described the five steps the Census used: harmonize variables across surveys, train a model, apply predictions to the larger survey, calibrate predictions to trusted benchmarks and aggregate to produce small‑area estimates. The agency tested multiple approaches and settled on an extreme gradient boosting ensemble, using an 80/20 training/test split and oversampling to address the relative rarity of households without air conditioning.

To align predictions with established measures, the Census calibrated results to higher‑level benchmarks including Energy Information Administration (EIA) and REC survey state estimates and adjusted for conceptual differences between whether a household "has" air conditioning and whether it "uses" it. For the Community Resilience Estimates (CRE) for heat, the Census combined social vulnerability components with heat‑exposure metrics drawn from FEMA (areas with maximum air temperature of 90°F or higher for two or more days in 2022) and National Weather Service measures of wet‑bulb globe temperature (80°F or higher at any time in 2022).

Sawyer said LACE produces tract‑level heterogeneity that county or state averages can mask. As an example, he said Salt Lake County matches the national AHS average at the county level — about 7% without air conditioning — but LACE shows older downtown and mountain‑area tracts with substantially higher shares without cooling than newer western neighborhoods.

During a question‑and‑answer period led by Gretchen Gooding, participants asked about statistical uncertainty. Sawyer said the Census currently "utilize[s] the replicate weights from the American Community Survey" in its uncertainty calculations and that the calibration step mitigates some machine‑learning error, but "the LACE estimates don't have a worked in measure of uncertainty around the idea of the machine learning model" and the team is seeking feedback from the experimental data‑product community.

Sawyer also said the Census intends to publish methodological materials and a Python package to let others run cross‑survey models on their data, but the agency is finalizing clearance steps to ensure no internal data are published accidentally. He said publishing predicted probabilities as a PUMS (Public Use Microdata Sample) feature is under consideration as a future add‑on.

The Census encouraged users to review LACE by county and tract and to send feedback about local plausibility to the program email provided during the session. The agency framed LACE as an experimental product that improves geographic detail on air‑conditioning prevalence without adding a new question to the ACS.