Proxy Temperatures Warp Climate Damages at the County Scale

Climate-impact studies that fill gaps between weather stations with proxy temperatures can understate damages by at least 15%, a new economics paper finds. Chu (Alex) Yu, an assistant professor of economics at Wake Forest University, and Richard T. Carson, a distinguished professor of economics at the University of California, San Diego, treated U.S. weather-station readings as ground truth and asked how well the usual stand-ins recover them.

The misses are large, they last across days, and they cluster in the hot and cold tails that set off heat alerts, crime spikes, and crop warnings. Yu and Carson’s public summary, issued September 16, 2026, says those proxies understate climate impacts by at least 15% on average. The tables underneath that average are less tidy: once heat and crime are rolled up to county and monthly files, the same relationship can shrink or more than triple.

How Climate Studies Fill Gaps Between Weather Stations

Thousands of papers estimate how temperature moves crime, illness, power demand, and harvests, but a thermometer is rarely sitting on the block, the ward, or the field those papers claim to measure. Stations stand miles apart, so researchers borrow a nearby reading or a smoothed map and treat that number as the day’s heat.

Yu and Carson, in the accepted temperature-proxy problem in climate-impact estimates, call that borrowed number a proxy and then test it the blunt way. They take daily mean temperature at U.S. stations as the truth, predict the same days with the proxies empirical work actually uses, and read the gap as measurement error in a relationship that would otherwise be clean.

The usual working assumption in this literature is that those gaps are small, random, and normally distributed, which would nudge damage numbers a little toward zero and then be ignored. The accepted manuscript says the opposite: the errors have strong spatial and temporal patterns, they are often large, and overestimates are common even though the average pull is downward. The worst of it sits in the tails of the temperature distribution, which is the part a heat plan actually uses.

THREE COMMON PROXIES

• Nearest station: Assign each location the reading from the closest monitor, a simple rule that still fails when terrain or distance intervenes.
• Inverse distance weighting: Blend several nearby monitors and give closer ones more weight, which is the paper’s city-day benchmark when it is built at the station itself.
• PRISM grids: A model that regresses climate on elevation and terrain, then fills a map cell even where no instrument stands.

The authors walk those constructions through four stations chosen for different climates and terrain: San Francisco, Boulder, Chicago, and Palm Beach. They also compare traditional interpolation with PRISM across the station network, then carry the same proxies into a crime file to see whether the statistical damage shows up in a real outcome.

The 15 Percent Average Hides a Threefold Swing

The crime test is built from incident-level reports in the FBI’s National Incident-Based Reporting System and companion state systems, aggregated to a daily panel of more than 5,000 cities in the contiguous United States from 2010 through 2019. Violent crime in that file is force or the threat of force, including aggravated assault, robbery, rape, and homicide. The working sample is 20,155,388 city-day observations, with crime counted per 100,000 people, and city-day inverse-distance temperature treated as the benchmark truth.

Wake Forest’s summary of the work says that when the same crimes and temperatures are rolled from city to county, the estimated effect of extreme heat on violent crime gets smaller. The property-crime tables in the manuscript show why a single county story is not enough. On days above 90°F, the city-day inverse-distance estimate is 0.28 additional property crimes per 100,000. At county-day the inverse-distance figure falls to 0.21, a 25% drop. At county-month the estimate rises to 0.86 to 0.93, more than three times the city-day benchmark.

HEAT AND PROPERTY CRIME ABOVE 90°F

Those county-month figures are not a rounding quirk. The Association of Environmental and Resource Economists, which publishes the journal, put the crime application’s span in a single line when it flagged the paper on August 7, 2026: biases from 30% too low to 300% too high, with fixed effects able to make the bias worse.

🌡️ New #JAERE research asks: how well do common temperature proxies (nearest monitor, PRISM) match true station readings?
Not well — the resulting biases range from -30% to +300% in a temperature–crime application, and fixed effects can make the bias worse
📉 #JaereFigureFriday pic.twitter.com/JkkcwAnGzU

— AERE (@AereOrg) August 7, 2026

A separate exercise in the paper, which recovers a known temperature-to-temperature slope instead of a crime slope, puts average downward bias at about 2% for traditional interpolation and about 6% for PRISM. That 6% is not the same quantity as the 15% understatement in the university summary, and neither number is the threefold county-month swing. Each test answers a different question. The planning problem is that governments do not get to pick the tidy average. They inherit whichever cut of the file their statutes and grant programs already use.

Boulder’s 15-Degree Blind Spot

Error is not sprinkled evenly across the map. In mountainous terrain such as Boulder, Colorado, standard weather proxies differed from actual station temperatures by an average of nearly 15°F. In flatter, densely monitored ground such as Chicago, the same class of error fell as low as 1°F. The manuscript says state-level average errors are largest in the Rocky Mountain and West Coast states and smallest in the Southeast and the Midwest, and that winter months, when temperature swings harder, make the gaps worse.

For one commonly used proxy, the average miss was 3.3°F, or 1.85°C. The Paris Agreement asks parties to hold warming well below 2°C above pre-industrial levels and to limit the increase to 1.5 degrees Celsius. That 1.5°C cap is 2.7°F. The typical proxy error in the study is larger than the international target the same research literature is trying to price.

WHERE THE PROXY MISSES

• Boulder, Colorado: nearly 15°F average gap between proxies and station readings in mountain terrain.
• Chicago: errors as low as 1°F on flat ground with a denser monitor network.
• One common proxy: 3.3°F (1.85°C) average error, above the Paris 1.5°C aim of 2.7°F.
• Slope test: about 6% average downward bias for PRISM against about 2% for traditional interpolation.

Traditional interpolation gets noisier as the nearest station recedes, especially beyond 100 km. PRISM holds a more stable error as distance grows, because it folds in elevation and terrain, but stability is not the same as a small bias. The public argument about weather stations usually stops at siting, the charge that a sensor in a parking lot or beside a runway is already corrupted. Yu and Carson start after that fight. They take the station reading as true and still find that the gap between stations is wide enough to move damage estimates by more than the Paris 1.5°C line.

County Files Do Not Preserve the Daily Heat Curve

Most U.S. heat policy is not written on a city-day file with inverse-distance temperatures. Hazard mitigation plans, crop programs, and many crime and health analyses run at county or state scale, often on monthly weather. As of 2023, only seven U.S. states had dedicated heat sections in the hazard mitigation plans FEMA requires, according to a Federation of American Scientists briefing by Vivek Shandas, Grace Wickerson, and Autumn Burton.

That briefing also notes that NOAA’s Heat Watch campaigns, run with the National Integrated Heat Health Information System, have found intra-regional temperature gaps of upwards of 20°F when volunteers walk sensors through neighborhoods. Those campaigns are mostly urban. They do not capture rural heat or the mountain microclimates where Yu and Carson find the largest proxy misses. A county file that averages a valley floor with a ridgeline is doing, in public data, what Heat Watch shows on foot: it is blending exposures that people do not actually share.

The manuscript is not a critique of any one crime paper. It uses crime because the heat-crime link is known to be nonlinear and large enough to see, which makes it a hard test for whether aggregation-induced error survives a standard regression. City-day inverse-distance temperature is the authors’ stand-in for truth. Every coarser cut is a file a county office could actually hold. The 25% fade at county-day and the more-than-threefold jump at county-month are what that office would publish without ever changing the underlying incidents.

If a heat-crime coefficient shrinks when the geography gets coarser, a county that sizes cooling centers, overtime, or hospital surge off that coefficient is planning for a milder problem than the daily city file describes. If the same coefficient more than triples after monthly pooling, the plan can overshoot on the hot tail while still missing the daily timing of the days that produced it. Either way, the number on the briefing slide is not the number the thermometers recorded.

A Nearby Station Can Beat the Gridded Map

PRISM, built by Oregon State University’s PRISM Climate Group, is the USDA’s official climatological data for the United States and a default input in U.S. climate-impact work. Christopher Daly, who wrote the mapping system, has described daily operations that ingest about 10,000 temperature stations, then estimate each grid cell from 15 to 25 nearby sites weighted by distance, elevation, coastal influence, and terrain. The freely posted product is a set of 4 km daily temperature grids for the United States.

Yu and Carson’s result is that this sophistication does not automatically win. Where the monitor network is dense, as in many urban settings, a nearest-station reading often carries less measurement error than the gridded product. PRISM’s average downward bias in their slope test is about three times the traditional interpolation average, even though PRISM’s errors vary less from place to place. For the nearest-monitor proxy with interpolation error, 44% of ground-truth locations produced a slope above 1, meaning the impact estimate ran hot rather than fading toward zero. After adding the extra error from averaging daily minimum and maximum to get a mean, that share was still 30%. PRISM produced a slope above 1 at only 6% of locations.

That last comparison is why some researchers treat PRISM as the conservative choice: it is more likely to understate than to overstate. Conservatism of that kind is still a miss, and it is a miss that grows in the fixed-effect specifications this literature prefers. In sparse or complex terrain, the gridded map can still be the better tool because a single nearby station may sit on the wrong slope or in the wrong valley. The paper’s rule is local. No proxy is best everywhere.

Put the Thermometers Where the Impacts Are

Yu and Carson do not stop at the diagnosis. The practical guidance is to match the temperature source to the network that actually exists, and to put more stations where people, farms, and businesses take the heat.

WHAT THE PAPER TELLS RESEARCHERS TO DO

• Dense urban networks: Try the nearest station before assuming a gridded product is cleaner, because extra smoothing can add bias.
• Sparse or mountain terrain: A physiographic grid such as PRISM can beat a single distant monitor, though it still will not erase tail error.
• The outcome’s location: Measure temperature at the city, field, or block where the crime, the illness, or the yield is recorded, not at the county centroid.
• The station network itself: Expanding and keeping dense monitors cuts the error at the source, before any regression is run.

Yu put the policy stake in one sentence when Wake Forest posted the study.

If climate-impact studies misestimate how severely extreme heat drives crime rates, crop loss, or health emergencies, policymakers may also misestimate the scale of the responses needed to address those impacts.

Chu (Alex) Yu, Assistant Professor of Economics, Wake Forest University

PRISM already pulls in about 10,000 temperature stations a day. The remaining miss is geographic: instruments still sit far from the fields, blocks, and ridgelines the damage papers claim to measure. Until the network is denser in those places, a county heat plan will keep reading a different temperature than the one people actually felt.

Frequently Asked Questions

What Is PRISM Temperature Data?

PRISM is a statistical mapping system from Oregon State University’s PRISM Climate Group that turns station readings into continuous grids of temperature, precipitation, and dew point. Thirty-year normals are issued at about 800 meters; the free daily and monthly files used in most empirical work are 4 km. The product is not homogenized, so Daly has warned it should not be used for century-long trend analysis, and monthly grids may not match the average of that month’s dailies because the two products draw on different station mixes.

Why Do Climate Studies Use Temperature Proxies?

Physical weather stations are miles apart, and much of the historical record reports only daily minimum and maximum, not an hourly mean. Yu’s 2020 UC San Diego dissertation found that replacing a true hourly mean with the average of the day’s min and max, by itself, can bias an economic temperature effect by 5 to 10%, with the sign of the bias depending on location and season. Proxies exist because the outcome data are almost never collected at a thermometer.

Are All Climate Damages 15 Percent Too Low?

No. The 15% figure is the authors’ summary of average understatement across typical impact work. In the paper’s own slope test, traditional interpolation averaged about 2% downward and PRISM about 6%, while 44% of nearest-monitor locations produced overestimates. The crime application then moved in both directions, including a county-month hot-tail estimate more than three times the city-day benchmark. The honest reading is that the sign and the size of the miss depend on place, season, proxy, and the scale of the file.

How Does the Paris 1.5°C Limit Compare With Proxy Error?

The Paris Agreement, adopted December 12, 2015 and in force from November 4, 2016, sets a dual temperature aim: well below 2°C, with efforts to hold the rise to 1.5°C. Converted, 1.5°C is 2.7°F. The 3.3°F (1.85°C) average error Yu and Carson report for one common proxy is larger than that 1.5°C aim, which means a routine measurement gap in U.S. impact studies exceeds the international target those studies are often used to value.