Detailed map of US air-conditioning usage shows who can beat the heat — and who can’t

(Press-News.org) LAWRENCE — As climate change produces ever more heat waves, how many homes in the U.S. lack adequate cooling? Who’s most vulnerable to lethal temperatures, and exactly where do they live?

A researcher at the University of Kansas has produced the most comprehensive and detailed map of air conditioning usage in the United States — data that can improve the understanding of AC access for public health officials, urban planners, emergency managers, economists, energy auditors, providers of social services, private industry and other stakeholders.

The research appears today in the peer-reviewed journal Scientific Data.

“This paper was about developing a nationwide air conditioning dataset,” said lead author Yoonjung Ahn, assistant professor of geography & atmospheric science at KU. “In the past, many studies have tried to estimate air conditioning ownership, but available public datasets were too limited. For example, the American Housing Survey provides data only at broad geographic levels, such as counties or metropolitan areas, and only for surveyed locations. Other sources, like the Energy Information Administration, collect small samples rather than data for the full population.”

Ahn uses spatial modeling, spatial analysis, spatial statistics and big data to research disproportional impacts of environmental hazards, like heat waves.

“This dataset is important for understanding how people experience extreme heat as the climate warms,” she said. “Even though 90% of Americans have some kind of AC, not all systems are equally effective. Portable or evaporative units, for example, don’t cool homes well in humid regions.”

Until now, detailed data on types of AC in use by people across the country hasn’t been available.

“That makes it difficult to assess who’s most vulnerable to heat, especially in rural areas where people may have fewer options to cool off,” Ahn said. “In urban areas, people might go to a library or pool, but in rural areas, people may work outside and come home to hot indoor environments.”

The KU researcher said her dataset could help public health officials and policymakers identify regions that lack adequate cooling and where people might need support programs or energy subsidies.

“It could also inform energy efficiency decisions, helping people choose the most appropriate cooling systems for their climate without unnecessary cost,” she said.

To compile and analyze more detailed information about AC use nationwide, the KU researcher used Dewey’s comprehensive real estate dataset at the household level, combined with variables she’d identified in earlier studies — such as housing type, building age, renovation year, race and ethnicity, historical housing policies and climate data.

Then, to paint a fuller picture of access to air conditioning, Ahn and co-author Christopher Uejio of Florida State University relied on machine-learning algorithms to fill gaps in the data.

“Because the Dewey data had missing information that wasn’t random, I used a random forest algorithm to impute missing values,” she said. “Some variables, like housing type, had too much missing data and were excluded. Then I used another machine learning model, XGBoost, to classify homes into four AC types: central AC, other types like window or portable units, evaporative coolers and none. Using those predictions, I created counts of AC ownership by census tract. XGBoost worked well because it handles both categorical and continuous data, with overall accuracy between 97–99. Accuracy varied by air conditioning type, ranging from 87% to 97%.”

Ahn said she was surprised by some findings.

“Previous studies only focused on metropolitan areas, but my dataset allowed me to compare rural and urban regions,” Ahn said. “There were clear differences. For instance, central and evaporative coolers were more common in rural parts of Oregon, while urban areas had more central and ‘other’ types or no AC at all. Moreover, approximately 20% of households in Florida used other types of air conditioning, while more than 95% of homes in urban areas had central AC.”

The KU researcher also found socioeconomic and demographic differences in the more detailed data, which were impossible to glean from data with less resolution.

“Climate and heating type were the strongest predictors overall, but the proportion of Hispanic residents was an important variable for some AC types, especially evaporative coolers and other types,” Ahn said. “Those units are concentrated in regions like California and New Mexico, which also have high Hispanic populations.”

Ahn said New York City didn’t align well with the model’s predictions, likely due to its unique characteristics compared to other parts of the United States. “The housing stock there is older, and incomes are high, but AC types vary in ways the model couldn’t fully capture,” Ahn said.

She acknowledged limitations in her data, which could motivate follow-up research.

“Some regions have higher missing data rates, which adds uncertainty,” Ahn said. “Metropolitan areas like New York City had especially high missing-ness, making those predictions less reliable. Another limitation is that my dataset represents current conditions, not historical trends. I hope to develop a historical dataset from 1980 onward, the last time the Census Bureau conducted a national AC survey.”

Finally, self-reported data from surveys, such as the American Housing Survey, may not accurately reflect actual usage, Ahn said. For example, people might report on such a survey that they don’t have AC, but they actually use a swamp cooler or portable unit.

In the future, Ahn plans to combine data from multiple sources to compensate for their limitations.

The researcher, who joined KU faculty in 2023, is planning follow-up research to include historical data and to work with students on future projects as her lab grows at KU. The research was supported by the National Academy of Sciences Gulf Research Program (SCON-10000677) and the University of Kansas General Research Fund (2302047).

Datasets are available here.

END