Too hot to work: Somanathan, Somanathan, Sudarshan & Tewari (2021), The Impact of Temperature on Productivity and Labor Supply

Across the world, years that run hotter than usual tend to be years of lower economic output, and the pattern is strongest in hot, poor countries (Dell, Jones and Olken 2012; Burke, Hsiang and Miguel 2015). Crops suffer, but so do sectors far from the farm, including industry.

Several explanations compete: heat stress on workers, more illness and death, more conflict, more natural disasters. They work on very different clocks, and that difference is what lets the authors tell them apart.

The authors’ strategy follows from this. If heat stress matters, it should appear in a worker’s daily output, in a factory’s annual accounts and in a district’s GDP, with effects of similar size. Before reading on, make a prediction.

Bottom line. Heat stress is the one climate channel that works within hours, so it can be measured directly with daily data. The paper argues it is also large enough to matter for whole economies.

Work generates heat, and the body must get rid of it to keep its core temperature stable. How easily it can do so depends mostly on the air temperature, but humidity and wind matter too, because sweat cools only if it can evaporate. When heat can’t escape fast enough, the only safe response is to work less intensely.

Laboratory researchers therefore often measure heat with the wet-bulb temperature, which folds temperature and humidity into a single number. Pooling lab studies, Hsiang (2010) finds that once the wet-bulb temperature passes about 25°C, task efficiency falls by roughly 1–2% for every extra degree. That is not an extreme threshold: it corresponds to about 31°C air at 65% humidity, a routine afternoon in much of India and well within what occupational-safety rules treat as safe.

Outside the lab, humidity records are patchy, so the paper measures heat with each day’s maximum temperature, which usually falls within working hours. (An appendix repeats the worker analysis with wet-bulb temperatures and finds the same patterns, with slightly larger effects.)

Lab evidence is only a benchmark. On a real shop floor, output also depends on wage contracts, supervision, machinery and incentives, and workers usually operate well below their physical limits. That is why evidence from real workplaces matters.

1. 1915

   Ellsworth Huntington relates day-to-day temperature swings to the output of workers and students.

2. 1946

   Mackworth finds that wireless telegraph operators make more mistakes in the heat.

3. 1989

   An international (ISO) standard sets out how to estimate heat stress on working people using a wet-bulb globe index.

4. 2010

   Hsiang’s review of lab studies: efficiency falls about 1–2% per degree above a 25°C wet-bulb temperature.

5. 2014

   Graff Zivin and Neidell show that US workers in heat-exposed industries spend less time at work on very hot days.

6. 2020

   Adhvaryu, Kala and Nyshadham find that cooler-running LED lighting on factory floors raises worker productivity.

7. 2021

   This paper brings factory-floor and nationwide evidence from Indian manufacturing.

Bottom line. Physiology predicts a fast, measurable loss of productivity once conditions turn hot and humid. The open question was how big that loss is in real workplaces, and whether it adds up to something visible in national output.

The study stacks evidence at three levels: individual workers or work teams, manufacturing plants, and districts. A survey of diamond firms adds a window on how businesses use air-conditioning. Every dataset is matched to daily weather.

Weather comes from two sources. Near the weaving and garment factories, the authors use readings from public weather stations. For the steel mill, which had no station nearby, and for the national analyses, they use the India Meteorological Department’s gridded data (1° × 1°), averaged over points within 50 km of Bhilai or over each district’s area.

Each day is then classified by its maximum temperature. For the national factory panel, the bins below are used throughout, so a whole year boils down to five numbers: how many days fell in each bin. The same colours mark these bins everywhere in this guide.

Official Indian statistics follow the April–March financial year; the paper labels each year by the calendar year in which it starts.

Bottom line. The microdata reveal the mechanism: what happens to people on hot days. The national panels reveal the scale: whether it adds up across the economy.

Heat can cut output in two ways: people get less done while at work, and they are more likely to stay away. The authors first look at weekly output, which blends both, and then use daily records to pull the two apart.

Bin cut-offs differ by site because climates differ. The Delhi-area garment plants use ten bins; at warmer sites the coolest bins are merged, so Surat’s lowest bin runs up to 29°C and the steel mill’s and southern garment plants’ up to 27°C. Everywhere, the top bin is above 35°C.

Without climate control, output falls in weeks with more hot days. With it, the decline disappears. In the steel mill, where machine operators sit in air-conditioned cabins, output even rises slightly on hotter days, perhaps because cooling is switched on only when it’s hot, or because cold weather makes metal set too quickly and causes defects.

Today’s heat versus last week’s heat

With daily data, the authors add a second set of variables: how many of the previous six days fell into the hot bins. That shows whether heat lingers through fatigue or illness. To keep the estimates precise, the cooler bins are merged, leaving two hot bins, 33–35°C and above 35°C, compared with days at or below 33°C.

Who stays home?

The same daily model, with absences as the outcome, shows how heat affects attendance.

Two patterns stand out. Absences rise after hot spells even in climate-controlled plants: cooling the shop floor doesn’t cool anyone’s home. And in those plants, only paid leave responds to heat. When a day off costs money, people tend to come in anyway.

For Surat’s weavers, both channels can be put in the same units. On days they work, weavers average about 134 meters of cloth; because attendance is patchy, the average over all days is about 51 meters.

Bottom line. At work, heat lowers productivity unless the workplace is cooled. Outside work, hot spells raise absenteeism, which workplace cooling can’t fix. For Surat’s weavers, about a third of the loss from a hot week comes through absences.

A handful of firms could be unusual. So the authors turn to the Annual Survey of Industries, follow Indian plants for 15 years, and ask whether they produce less in years with more hot days. The model mirrors the worker model, one level up.

Every step down is a loss. A single day moving from below 20°C to above 35°C costs a plant about 0.22% of its annual output. That sounds small until a year’s worth of hot days is added up.

Warming scenarios

Adding state-specific quadratic trends, or controls for floods, industrial conflict and power cuts, barely moves the estimates.

Same answer, different ruler

Counting days in bins is one way to summarise a year’s weather. The authors also try degree-days, which record how far each day’s temperature climbs through each bin, and models in which output depends on the sum of daily temperatures and their squares.

All versions point the same way. The polynomial versions give somewhat smaller point estimates, but their confidence intervals overlap with the others.

Bottom line. The factory-floor pattern holds across India’s factory sector: more hot days, less output. The result survives every alternative specification the authors try.

Why would a hot year shrink a factory’s output? The authors write down a Cobb-Douglas production function and let temperature shift three things: overall productivity, and the output elasticities of labor and capital, which measure how much output rises when an input rises by 1%.

• ↓

  Labor’s elasticity falls with heat. Its interaction with temperature is negative in all four hotter bins and statistically significant. Each extra day in a hotter bin, instead of below 20°C, lowers it by roughly 0.0008 to 0.0013.

• ↑

  Capital’s elasticity edges up. The capital–temperature interactions are small and positive.

• 0

  No leftover productivity effect. Once the labor and capital interactions are included, temperature has no significant direct effect.

Could firms simply adjust their workforce after seeing the weather? Within a year that seems unlikely. Capital is measured at the start of the year, and India’s labor laws are rigid: the Industrial Disputes Act of 1947 requires government approval before firms with more than 100 employees dismiss workers, and in 2017 the World Bank ranked India 130th on ease of doing business, citing such rules. Still, as a check the authors use the Levinsohn–Petrin estimator, which lets labor respond to shocks. The labor–temperature interactions come out smaller but still negative, and statistically indistinguishable from the main results.

Cooling could break the link between heat and labor. But during the study period, Indian factories had very little of it.

Bottom line. Hot years cut the payoff from labor, and the plants that rely most on workers lose the most. The problem isn’t the machines; it’s people working in uncooled heat.

To line up with cross-country studies, every estimate is translated into the same thought experiment: every day of the year becomes 1°C warmer. Step through the levels, from a sewing line to whole economies.

The estimates are strikingly similar across scales, and close to what cross-country studies find for industry in poor countries and for total output.

That doesn’t show that heat stress causes all of the country-level losses. But those studies use data going back to the 1950s, when cooling was rare nearly everywhere. If workers then responded to heat the way Indian factory workers do, heat stress alone would be enough to explain the entire effect.

Bottom line. Heat stress on workers is large enough to account for country-level temperature losses, which makes it a far more important channel than previously believed.

Cooling works: in the garment firm, climate-controlled lines show no productivity loss on hot days. But it costs electricity. For Surat’s weaving units, a back-of-the-envelope calculation in the paper’s appendix finds that air-conditioning would cost a lot relative to the output it would save, because value added per worker is low. The garment firm, where each worker adds more value, had cooled most but not all of its lines.

If heat hurts mainly through labor, firms that do invest in cooling should target it at tasks that rely on workers and matter most for the value of the product. Diamond processing in Surat is a good test. Units there run five operations (sorting and grading, planning and marking, bruting, which rounds the stone, cutting, and polishing), some largely mechanized and others heavily hands-on.

There are also signs of adaptation over time. In the factory panel, the damage from a day in the two hottest bins shrinks by about 6–8% per year over the 15 years, and plants that use more electricity per unit of output, as cooled plants would, respond less to heat. As economies grow richer, manufacturing may become less exposed.

Bottom line. Adaptation is happening, but slowly and selectively. Cooling pays where labor adds a lot of value, and tends not to pay in low-value, labor-intensive work, which is exactly where heat bites hardest.

Heat could lower factory output without touching anyone’s productivity. The authors test the main suspects. Open each one to see the evidence and the verdict.

Bottom line. None of the alternatives explains away the temperature effect. Lower labor productivity remains the best explanation.

Longer-run responses, such as more automation, relocation, and a shift away from labor-intensive production in hot places, could widen wage gaps if they favour more productive workers. The logic also reaches beyond factories: construction and farm work are hard to cool, so some crop losses usually blamed on plant biology may partly reflect heat-stressed farm workers. The authors leave these as open questions for future research.

Annual Survey of Industries (ASI)

India’s main survey of registered factories: every large plant plus a random sample of smaller ones.

Cobb-Douglas production function

A standard way to model output as a product of inputs, each raised to a power; in logs it becomes a simple sum.

Conley standard errors

Standard errors that allow outcomes in nearby places, and in the same place over time, to be correlated. Here, correlation is allowed within 150 km.

Degree-days

How far a day’s temperature climbs through a given temperature range, summed over the year.

Fixed effects

Controls that absorb anything constant for a unit (a worker’s skill, a plant’s location) or common to a period (a nationwide shock), so estimates come from changes within units over time.

Inverse hyperbolic sine

A transformation that behaves like a logarithm for large values but, unlike a log, is defined at zero, which matters when a worker produces nothing on a day off.

Lagged temperature

In the daily worker models, the number of hot days among the six days before the current one.

Levinsohn–Petrin estimator

A production-function method that uses spending on intermediate inputs to control for productivity shocks researchers can’t see, allowing inputs such as labor to respond to them.

Omitted (reference) bin

The coolest temperature bin, left out because the day counts always add up to the total number of days. Every bin coefficient is measured relative to it.

Output elasticity

The % change in output when one input rises by 1%, holding others fixed.

RCP 8.5

A high-emissions greenhouse-gas pathway used in climate projections.

Temperature bins

Ranges of daily maximum temperature. Counting days per bin lets the relationship between temperature and output take any shape.

Wet-bulb temperature

A measure of heat that combines temperature and humidity; it reflects how well sweating can cool the body.

This guide paraphrases the published paper and redraws its results. Numbers come from the paper’s text and tables, except where marked as an illustration or as approximate values read from a figure.