When economists ask how temperature affects the economy, they usually pull out annual or quarterly data. That framing has fed a long-running debate: rich, temperate countries like the United States often appear to shrug off heat, while poorer, hotter economies seem to bear the brunt. But averaging economic activity over months or years can smooth away exactly the short-run wobbles that matter most when the thermometer climbs.
A new NBER working paper zooms in on weekly data to trace how temperature shocks ripple through state-level U.S. economies over the following two years. The authors find that a hotter-than-normal week produces a small early bump in economic activity, followed by a delayed and persistent slump that plays out primarily through the labor market.
Rethinking the question with higher-frequency data
Kimberly A. Berg of Miami University, Chadwick C. Curtis of the University of Richmond, and Nelson C. Mark of the University of Notre Dame set out to test whether the muted temperature effects seen in annual U.S. data are real, or whether they are an artifact of low-frequency measurement. Their reasoning: both temperature and economic activity move quickly, and averaging them into big time buckets could hide meaningful dynamics.
To capture those dynamics, the researchers combined two weekly datasets. For temperature, they used NOAA’s daily gridded temperature records, weighted by population within each state, and averaged them to a weekly figure. For the economy, they used the weekly state-level Economic Conditions Index (ECI) developed by Baumeister and colleagues in 2024. The ECI blends roughly 21 indicators per state, including unemployment claims, employment, GDP, retail gas prices, mobility, and card transactions, and it is scaled so that zero corresponds to the long-run U.S. growth rate.
The sample covered the continental United States from April 1987 through June 2024, a stretch that includes several recessions and the COVID-19 shock. To keep those national events from muddying the results, the analysis included fixed effects that soak up shocks affecting all states in any given week.
Tracing the response week by week
To follow the economic aftershocks of a temperature shift, the researchers used a statistical approach called panel local projections, which estimates the effect of a shock at each future horizon rather than assuming a rigid model of how things unfold. They ran the projections out to 104 weeks, or two years, and included a squared temperature term to allow for the possibility that the response bends at higher heat levels.
The baseline exercise asked: what happens to a state’s cumulative economic conditions after a one-degree-Celsius rise in weekly temperature, starting from the sample-average state temperature of about 12.6°C?
The answer came in three stages. In the first six weeks or so, economic conditions edged up modestly. After roughly ten weeks, the cumulative response flipped negative. From there, it kept sliding, bottoming out around 80 to 90 weeks after the shock before beginning to stabilize.
The near-term uptick was small; the medium-run decline was the dominant feature. Because the ECI is measured relative to long-run U.S. growth, this pattern implies that hotter weeks nudge states below their normal growth trajectory for well over a year.
Warmer climates take a bigger hit
The quadratic temperature term let the authors ask whether the same one-degree bump has different consequences in cool states versus warm ones. They plotted responses at base temperatures ranging from North Dakota’s chilly 5.5°C average to Florida’s 23°C.
For roughly the first year after a shock, states looked broadly similar. But at longer horizons, the responses fanned out. Warmer base climates kept sliding, while cooler and median-temperature states stabilized and started to recover. In other words, an extra degree in Florida accumulates into a larger and more lasting economic drag than the same extra degree in North Dakota.
The researchers also checked whether the result depended on the exact way they modeled nonlinearities, defined temperature anomalies, or seasonally adjusted the raw data. Across several alternative specifications, the delayed medium-run decline held up.
Season matters, but less than you might expect
A natural intuition is that summer heat should be doing most of the damage. To test that, the authors let the temperature response vary by season. Summer shocks did produce somewhat larger medium-run declines, and winter shocks somewhat smaller ones. But the differences were modest, and the same delayed-and-persistent pattern showed up in every season.
That finding pushes back gently against a story in which summer heatwaves are the sole culprit. Temperature fluctuations at other times of year appear to be feeding through the economy as well.
The labor market does most of the work
The ECI can be decomposed into six categories: labor market, mobility, real activity, expectations, financials, and households. When the authors ran the same analysis on each component, the labor market response stood out. It closely tracked the shape and timing of the overall decline and accounted for the largest share of the medium-run drop. Mobility showed a similar but smaller pattern; households declined more gradually; the other categories moved little.
To dig further, the researchers looked at three specific labor measures: monthly total nonfarm employment, monthly average weekly hours of manufacturing production workers, and weekly unemployment insurance initial claims.
Hours worked showed little immediate change, then declined starting a few months out and stayed depressed through the medium run. Employment barely moved, dipping slightly before drifting back toward zero. Weekly unemployment insurance claims initially declined a touch, then rose steadily starting around week 30 and continued accumulating.
The authors interpret this pattern as gradual labor market stress rather than sudden layoffs. Firms appear to trim hours and slow hiring before cutting headcount outright, while separations gradually accumulate in the claims data.
Building a model to explain the delay
One puzzle stands out in the results: weekly temperature is only weakly persistent in the data. A single hot week does not by itself last long. So why do its economic effects still show up a year and a half later?
To explore possible mechanisms, the authors built a weekly labor search and matching model with three added ingredients. First, higher temperatures raise the disutility of supplying labor hours, capturing fatigue and reduced worker effort. Second, temperature shocks feed into a “heat damage stock” that accumulates and decays slowly, converting brief weather events into extended economic pressure. Third, an “adaptation stock” builds up gradually and offsets some of the damage over time, reflecting slow adjustments like changing work schedules or investing in cooling.
Because adaptation moves more slowly than heat damage in the model, net heat exposure stays elevated for a stretch before adaptation catches up. That combination generated a persistent medium-run decline in the model’s simulated labor market index that broadly mirrored the empirical response, including a delayed trough and eventual recovery.
When the authors turned off individual pieces of the model, each played a distinct role. Removing hours adjustment costs made labor utilization drop faster on impact. Removing heat damage persistence eliminated the medium-run propagation. Removing adaptation eliminated the eventual recovery.
Implications and caveats
The researchers argue that the standard reading of rich, temperate economies as largely insulated from temperature shocks may partly reflect the frequency of the data used to study them. When measured week by week, U.S. states show clear signs of adjustment that annual figures can wash out.
Several caveats apply. The ECI is a composite index rather than a direct measure of output, and the local projections estimate associations rather than proving causation, though the use of week and state fixed effects helps address common shocks. The authors also note that their model is meant to illustrate plausible channels rather than deliver a full quantitative account of the data.
