I reviewed A Cold Stop: Temperature, Unemployment and Joblessness Dynamics by Graff Zivin, Lepinteur, Neidell, and Nieto Castro (NBER Working Paper No. 34487, November 2025).
The paper looks at how temperature shocks affect unemployment in the United States. The authors build a model of how firms and workers behave (hiring, firing, quitting, job searching) to figure out the different ways temperature could affect unemployment. They link high-frequency weather data to three decades of monthly Current Population Survey data with over 15 million observations. They find that cold temperatures increase unemployment, specifically in weather-exposed industries. Cold increases layoffs but not quits or temporary contract endings, and there are more absences and operational disruptions. The effect is demand-driven, meaning firms reduce hiring during cold weather.
I think this paper is clearly written and makes a meaningful contribution to climate and economics research. The research question is timely given current job market conditions, and the finding about cold affecting unemployment is new.
The Research Design
The way the authors isolate the effect of temperature, using variation within the same county, year, and state-by-month, works well for the question. They also run a lot of robustness checks which makes their findings stronger.
One thing that stood out is how the coefficient on the coldest temperature bin drops to zero when they control for snowfall. They say snowfall is probably a bad control since it depends on temperature, but it makes me wonder if they're really picking up on cold temperatures or snowy/icy conditions, which would have different implications. One way to dig into this would be to split regions that get a lot of snow from regions that just get dry when it's cold. That would help separate the effects.
The authors also observe temperature over the past three months as one chunk, but I think they should look at each month separately so we can see where the cold matters most. For example, does the cold affect unemployment right away, or does it take a couple months?
Also, the county assignment could be improved. They use the Current Population Survey, which tells you where people live, but for almost half the sample it just says "metro area" without specific counties. They assign people to the biggest county in the area, which could mean matching people to the wrong weather data. They discuss dropping those observations without the results changing, but it's still worth addressing since it's almost half the data.
The Framing
The motivation is strong. They clearly discuss why weather shocks disrupting employment matters for preparing for climate change. The theoretical model helps organize the different ways temperature could affect unemployment. I just think 4 pages on it might be excessive since the same information could be described more briefly.
The empirical sections are where the motivation really comes through, walking through each channel step by step. Their argument for why cold matters more than heat is convincing too. Cold weather creates hard cutoffs where work just can't happen, while heat mostly slows things down. That makes intuitive sense and helps explain why their results look the way they do.
The Evidence
The main CPS analysis is strong with over 15 million observations, a solid identification strategy, and robustness checks that all point in the same direction. The mechanisms section could be improved though, since the ATUS sample for job search is very small at about 5,600 observations and the result doesn't tell us much.
The overall conclusion that cold weather increases unemployment through the demand side holds up well, but the evidence for specific mechanisms is limited by the data.
Overall
There aren't fundamental problems and the main finding is new and policy-relevant with a credible research design. The most important revision would be digging deeper into whether the effect is really about cold temperatures or snow and ice. They could also break the high-risk industries into more specific categories since they could face different cold exposure. It should all be doable with their data.