## Discussion

In the face of a warming world, workers are already shifting schedules to limit midday heat exposure20,21,22,23, but daylight hours are limited26, and as shown here, background global warming will increasingly restrict the ability of workers to adapt to warming by time shifting (Fig. 1). Even at the coolest hour of the day, there are currently several billion hours of heavy labor lost per year globally (Supplementary Fig. 2), with labor losses in the early morning hours increasing nonlinearly as the globe warms (Fig. 3). The relationship among global temperatures and global total labor losses and economic productivity losses is inherently nonlinear as the background climate state changes and the geographic extent of heat exposure increases (Figs. 3, 4; Supplementary Fig. 8).

Labor losses from heat exposure are spatially variable, with several countries in Southwest Asia, South Asia, and Africa that already experience per-capita, 12-h workday labor losses > 200 h/person/year. Qatar and Bahrain show the worst impacts, with >300 h/person/year labor losses (Fig. 5a). In the coolest hour of the day, Qatar and Bahrain are still the most impacted by heat exposure (15 h/person/year lost), followed by several island and coastal nations in the western Pacific, which show losses >10 h/person/year at this coolest hour (Supplementary Fig. 11). When we overlay per-capita labor losses on the working-age population in heavy outdoor labor (Methods), we find that countries with large populations in South and East Asia experience the most work hours lost, both in the coolest hours (not shown) and in the full workday (Fig. 6a), with India showing the largest heat exposure impacts on heavy labor (>101 billion hours lost/year), despite its modest average per-capita labor losses (162 lost hours/person/year). Large population-weighted labor losses (>10 billion hours/year) in other countries such as Pakistan, Bangladesh, and China are driven by a combination of large working-age populations, seasonal heat exposure, and large fractions of the population that work in agriculture and construction industries (Supplementary Fig. 12). Under future warming, India, China, Pakistan, and Indonesia experience the largest population-weighted labor losses (Fig. 6b–d) and associated economic productivity impacts (Supplementary Fig. 13), despite having lower national-average per-capita losses than other countries with smaller populations in Southeast Asia and tropical Africa (Fig. 5b–d). Bangladesh is a notable exception as it shows large per-capita as well as population-weighted labor losses currently and with warming.

Our accounting assumes that individuals are losing work productivity in the heat. Indeed, laborers who are encouraged to self-pace may regulate their own workload to maintain comfort19. However, worker productivity is linked to economic incentives, which is in turn linked to the health and well-being of workers, so individuals may continue to work at the detriment to their health, such as when they are paid by the piece for work29,30,31. If laborers are unable to work under safe conditions, they are at higher risk of multiple health impacts, including premature death32,33,34,35, workplace injuries36, morbidity from heat-related illness37,38, traumatic injuries7,39, and acute kidney injury31. Heat exposure is also implicated as a potential contributing factor to an epidemic of chronic kidney disease of unknown etiology in otherwise healthy, relatively young workers in Central America, Sri Lanka, India, and Egypt, and other areas40,41. Heat exposure can also increase the absorption of certain chemicals42 and is associated with adverse pregnancy43 and mental health outcomes44.

We have focused on worker adaptation via moving work hours from the middle of the day to the early morning hours23,24,26; governments have already implemented mandatory work breaks during the hottest parts of the day in locations such as in the United Arab Emirates19. However, implementation of this approach is situationally dependent and can result in unintended consequences. First, moving work to earlier hours may impact sleep duration, which is associated with injury risk45,46. Furthermore, heat exposure can affect sleep47, which can affect the risk of injury and heat strain. Approaches to optimize sleep hygiene, and consideration of impacts on circadian rhythms and sleep, should be included in plans to shift work hours. Second, occupations and industries (e.g. construction) in certain settings may be limited in their abilities to shift work hours due to policies such as local noise ordinances48. However, policies that restrict night shift or early morning work, such as noise restrictions, may not be permanent barriers to adaptation if future investigation of potential adaptation strategies prompt changes in local ordinances to accommodate these strategies. Also, changing work hours has the potential to introduce additional hazards related to other aspects of ambient conditions, such as lighting. These factors should be anticipated and addressed when optimizing work hour timing. Finally, changes in work schedules need to be coordinated with childcare and other obligations to maintain overall community well-being. Workers and communities should be included in decision-making to ensure that important considerations are not overlooked. Nonetheless, our findings provide baseline climate information about shifting work times, which is critical to informed decision-making about the most promising combination of approaches at different levels (e.g., individual, workplace, community, policy).

Future work should also consider including other heat stress metrics17, because different metrics are best used for approximating heat stress in distinct situations, such as considering extent of perspiration and clothing49. We have also used reanalysis-based hourly estimates of heat exposure, but this method is known to be conservative as it underestimates extremes observed at weather stations1. Additionally, to better account for uncertainty in projections of climate impacts, future approaches could incorporate more detailed on-the-ground data related to environmental conditions, work-rest cycles, work pace, work organization, physiological heat strain50, and other ERFs that relate work capacity to heat exposure51. Although reanalysis-based global estimates of annual heavy labor losses due to heat exposure approach several hundred billion hours per year13 (Supplementary Fig. 2), there are relatively few field-based studies that quantify work time lost due to heat exposure5,23; more field observations are needed to verify the results presented here. Additionally, this study has focused on one specific adaptation mechanism, shifting of work to cooler hours. Future work could consider other adaptation strategies, such as task shifting (e.g., movement of labor-intensive tasks to cooler hours), the limits of time and task-shifting strategies, what these mechanisms will cost both together and separately in terms of lost productivity, as well as recommendations for how workers could make choices between adaptation options and weigh their utility under various warming levels and under different scenarios.

With global-mean temperatures now over 1 °C warmer than a century ago, the Earth’s climate is still within a regime that makes moving worker hours an approach – when combined with other adaptation mechanisms – that can help cope with warming temperatures. If future warming can be limited, this time-shifting adaptation mechanism may remain an effective option for many locations in the tropics and subtropics (Fig. 3, Fig. 4). However, with additional warming, this adaptation mechanism becomes less efficient as unsafe heat exposure in the morning hours magnifies in the tropics and subtropics, and expands into the extratropics. An additional 1 °C of global warming relative to the present could occur as early as 2037, and another 2 °C of warming could occur as early as 2051 (Supplementary Fig. 14). Therefore, if warming is left unchecked, the globe will continue to move into a new, ‘less adaptable’ climate regime within the lifetime of many young and middle-aged workers. These results further highlight the need to find alternative adaptation mechanisms to keep workers safe as well as to limit future warming to 1.5 to 2 °C52 to help protect the livelihoods and health and well-being of workers in the low and mid-latitudes.

## Methods

### Heat exposure, labor losses, and associated economic costs

There are a variety of methods for estimating heat exposure17,53,54. Wet Bulb Globe Temperature (WBGT)55 is an internationally recognized heat stress metric that incorporates temperature and humidity measures and is used in occupational health studies and in military applications56,57,58,59. The value of the WBGT metric (or similar metrics that account for heat and humidity) is that it enables us to determine heat stress for both dry and humid heat in a common way (i.e., how easily the human body is able to cool itself). The ability to compare risk in both dry and humid conditions is essential in adaptation planning because use of air temperatures alone would not take into account the differences in heat stress due to variations in relative humidity throughout the day, or across seasons or locations. However, WBGT has its own limitations60, and in some cases is not the best metric for measuring heat stress61,62,63. Additionally, WBGT can be difficult to measure and calculate as it requires specialized equipment that is not typically used at weather stations56, and the necessary measurements needed to estimate WBGT are not output as variables from state-of-the-art model projections. Therefore, we focus on sWBGT to estimate heat exposure and labor impacts. sWBGT approximates WBGT using estimates of near-surface temperature, humidity, and pressure. The sWBGT metric used here assumes no solar radiation and is intended to estimate heat exposure in the shade or indoors with no air conditioning1. It is important to note that WBGT in the sun can be at least 2–3 °C higher than shade values64, and reanalysis-based estimates of WBGT can underestimate extremes1, so our estimates of productivity losses from heat exposure may be conservative. Nonetheless, these shade values of WBGT are used to estimate warming impacts on labor in the ILO report on labor on a warmer planet11, in the Lancet Countdown on Health and Climate Change13,65,66, and in other recent work1. Further details about calculation of sWBGT in reanalysis data can be found in the sections below.

We use an exposure response relationship based on epidemiological data to derive ‘work ability’ (WA) for a given hourly value of WBGT. This method13,65,66 employs estimated exposure response relationships for reduced hourly work capacity (labor productivity) for heavy manual labor conducted at 400 W intensity using hourly sWBGT and a cumulative normal (ERF) function:

$${{{{{\rm{Loss}}}}}}\,{{{{{\rm{fraction}}}}}}=1/2\times (1+{{{{{\rm{ERF}}}}}}\,({{{{{\rm{sWBGT}}}}}}-{{{{{\rm{WBGTaver}}}}}})/{{{{{\rm{WBGT}}}}}}{{{{{\rm{SD}}}}}}\times \surd 2)\left)\right.$$
(1)

where for heavy work WBGTaver = 32.47 and WBGTSD = 4.16 for productivity lost per person per hour. Previous work25,65,66 has assumed labor loss cutoffs of 10% and 90% of the hour, but here we use the approach employed in the 2020 Lancet Countdown on Health and Climate Change13 that assumes the amount of work loss is defined by the exposure function. Although heat exposure can impact workers conducting light (e.g., services) and medium (e.g., manufacturing) labor, here we just consider heavy labor impacts because heavy labor losses account for the largest fraction of labor loss due to heat exposure11,13,66.

We examine work losses in the 12-h workday, in the hottest hour of the day, in the third coolest hour of the day (here referred to as ‘morning hour’), and in the coolest hour of the day (typically close to sunrise around 5–7 a.m.  in the warm season). We also investigate the amount of time that could be recovered by moving labor from the three hottest hours of the day to the three coolest hours of the day, assuming daylight is not a limiting factor. To estimate labor losses during the 12-h workday, we calculate the daily mean sWBGT, daily maximum sWBGT, and the halfway point between these two values25, and assume 4 h is spent near each of these values in the 12-h workday (4 × sWBGT max + 4 × sWBGT mean + 4 × sWBGT half). Although hourly weather reanalysis data are now available to calculate hourly losses in the 12-h workday, we have chosen to use the established ‘4 + 4 + 4’ method due to computation and data storage constraints and for better comparison with previously published results13; further discussion of this method can be found in Supplementary Text 1.

The theoretical annual maximum work loss per person in the 12-h workday is 4380 h/year (12 h/day, 365 days/year), and for the hottest hour, morning hour, and coolest hour, is 365 h/year (or 1 h/day, 365 days/year). Here we focus on the 12-h workday based on the idea that most workers conduct their work between approximately 7 a.m. and 7 p.m.13,25,28,66. However, this method does not account for unsafe heat exposure outside of the traditional daylight work hours, such as late evening hours, when heat exposure is still high in many low-latitude locations (Fig. 1); during these ‘non-traditional’ work hours, unpaid household labor is often still conducted67.

Our calculations likely provide conservative estimates of heat exposure for several reasons. Newly released, empirically based physical work capacity estimates51 indicate that current methods11,13 could underestimate work loss under heat and humidity levels previously thought to cause little to no productivity losses. Also, here we consider sWBGT calculated from ERA5 data, which, as previously mentioned, underestimates heat exposure extremes observed at weather stations1. In a location like the southeastern United States, sWBGT shows minimal to no labor productivity losses in the current average summer month (Fig. 1), but agricultural workers are already experiencing adverse heat health outcomes in many parts of the USA, so our estimates of labor losses underestimate actual worker risk21,32,68,69. The sWBGT metric used here considers heat exposure in the shade, so it will underestimate heat exposure in the full sun, and some work cannot be conducted in or moved to shaded areas. Finally, we do not take into account additional factors that could influence worker safety and productivity in the face of high heat and humidity such as ability of workers to use different clothing69, underlying health conditions, varying degrees of acclimatization to heat, or hydration, among other factors.

We use ILO70 estimates of numbers of workers in each country who work in heavy labor, here defined as agriculture+forestry+fishing and construction13 to quantify the number of heavy labor work hours lost. For each of the countries in the dataset that have relevant ILO data that overlapped with the World Bank data (n = 163), we use the fraction of the overall working-age population (ages 15–64) in that country that works in heavy labor, multiply this fraction by the spatially gridded population ages 15–64 (Gridded Population of the World v4 data71), and then overlay the hours lost on the population data13,25,65,66. This method assumes that outdoor workers are geographically distributed similarly to the overall population, even if this is not always the case69. Nonetheless, for most countries sub-national information on work is not available, so we follow established methodology that distributes laborers with the general population.

We also estimate economic productivity decreases associated with lost earnings from heavy labor productivity loss. There are several methods to estimate economic costs, including multiplying hours lost by estimates of hourly earnings13,72 and converting hours lost to job loss equivalents and associated productivity losses in terms of reduced contributions to Gross Domestic Product (GDP)11. We use the most recent World Bank GDP data to convert average productivity per worker in agriculture, forestry, fisheries and industry (including construction) to hourly output by assuming a 12-h workday, 365 days/year to maintain consistency with our hours lost estimates. We then multiply the hourly productivity per worker by the heavy labor hours lost to estimate economic costs of productivity losses due to heat exposure (details in Supplementary Text 1).

For future warming impacts on labor, we assume future population and earnings are static—in other words, they are fixed at levels and rates from the present. This is a conservative assumption for projecting future population impacts because future population is expected to rise, particularly in many low-latitude countries where heat exposure is projected to increase73,74.

### Heat exposure metric calculation from reanalysis data

Following the method of Li et al.1, we calculate the sWBGT from hourly, single level (near-surface) ERA575 atmospheric reanalysis output (Jan 1, 1979 to Dec 31, 2020) using 2-m air temperature (t2m), surface pressure (sp), and 2-m dew point temperature (d2m) using the equation:

$${{{{{\rm{sWBGT}}}}}}=0.7{{{{{\rm{Tw}}}}}}+0.3{{{{{\rm{Ta}}}}}}$$
(2)

where Tw is ‘isobaric wet bulb temperature’ and Ta is dry air temperature (t2m). Tw is calculated from air temperature, dew point temperature, and surface pressure. ERA5 is provided at ~35 km spatial resolution at the equator, and climate model data are regridded to this grid resolution via bilinear interpolation when the model-based warming patterns are added to the reanalysis data. When ERA5 sWBGT and hours lost data are compared to population data, the hours lost spatial data are regridded to the population resolution (~0.5 × 0.5 degree spatial resolution). We overlay a spatial mask of each country’s borders to calculate the country-by-country labor loss estimates by sector and worker productivity losses, and sum labor and productivity losses within the country borders to calculate country-level losses.

### Heat exposure metric calculation from climate model data

Following the methods of14, we calculate WBGT using 2-m air temperature (‘tas’ variable), near-surface specific humidity (‘huss’ variable), sea-level pressure (‘psl’ variable), and orography (‘orog’ variable) from CMIP6 models that provide the necessary variables (Supplementary Table 2) using the equation:

$${{{{{\rm{sWBGT}}}}}}=0.567T+0.393{VP}+3.94[{\,}^\circ {{{{{\rm{C}}}}}}]$$
(3)

where T is daily mean 2-m air temperature (‘tas’) and VP is vapor pressure. Vapor pressure (VP) is calculated from daily mean specific humidity (‘huss’), sea-level pressure (‘psl’), and orography (‘orog’). We use the output from idealized 1%CO2 simulations, which are only forced by increases in atmospheric CO2 concentrations, starting at pre-industrial levels (~284 ppm), and increasing at 1% per year for 150 years27. Here we use model years 35–150 from the 1%CO2 experiment because year 35 approximately coincides with present atmospheric CO2 concentrations (~400 ppm). We choose the 1%CO2 experiment because the 21st century Shared Socioeconomic Pathways (SSPs) include highly uncertain, theoretical future transient aerosol, land use, and other forcing changes73. To determine if warming patterns among experiments are robust, we compare warming patterns from 1%CO2 to patterns from the SSP5-8.576 and find that warming patterns are similar (<10% difference in local magnitude), except in isolated locations in the mid-latitude northern hemisphere (see Supplementary Fig. 15 and Supplementary Text 1).

### Warming patterns in CMIP6 models

We first calculate daily sWBGT for each CMIP6 model then calculate the monthly mean sWBGT for each grid point. We regress global-mean, annual-mean, latitude-area weighted 2-m air temperature (‘tas’) against monthly local sWBGT after smoothing global and local data using a 20-year lowpass filter. We use the multi-model median (MMM) regression coefficient from this calculation (e.g., local change in each month per degree of global change) as the ‘pattern scaling’ variable for different global-mean temperature changes examined here. We have also calculated the warming patterns for annual, JJA/DJF, the 75th percentile, the 95th percentile, and the 99th percentile of daily sWBGT and find minimal differences in MMM warming patterns for land regions between ~40° N and 40° S (Supplementary Fig. 6). Warming patterns are calculated on the model’s native grid resolution, then spatially regridded using bilinear interpolation to the ERA5 resolution for calculation of MMM and for adding warming patterns to the ERA5 data. We choose the MMM instead of multi-model mean because local warming patterns among CMIP6 models are not normally distributed at some locations. Supplementary Fig. 3 shows sWBGT warming patterns from individual CMIP6 models. To show where CMIP6 models agree on the magnitude of local change per degree of global warming, we calculate the coefficient of variation (inter-model local standard deviation divided by the local MMM). We stipple locations on maps (Fig. 2, Supplementary Figs. 47) where the coefficient of variation is > 0.35 to show where models do not indicate there is good agreement on the magnitude of local change relative to the MMM16.

### Applying warming patterns to reanalysis data to estimate future heat exposure

Here we are interested in quantifying heavy labor losses due to heat exposure, with a focus on heat exposure in the peak heat of the day (daily maximum), in the morning, and in the coolest hour of the day (typically around sunrise). We combine monthly sWBGT warming patterns from CMIP6 models with hourly sWBGT ERA5 data to estimate future heat exposure impacts on labor. This combination allows us to rely on the instrumental-based background climate mean state and available model data without the need to bias correct model data. For the present-day climate, we calculate hourly sWBGT in ERA5 data, apply the productivity loss equation, and then calculate the mean work hours lost for the time period 2001-2020. To calculate future productivity losses, we add the monthly warming patterns from CMIP6 models to the hourly ERA5 sWBGT data (e.g., for a global warming of 2 °C in January, we multiply 2 by the local January CMIP6 warming pattern, then add this number to the hourly January ERA5 sWBGT data). We use monthly warming patterns because warming patterns can vary by season (Supplementary Fig. 5). We conducted a sensitivity test using warming patterns from the 75th percentile of sWBGT and hemispheric summer warming patterns (JJA and DJF), and our main results did not change (not shown). After adding CMIP6 sWBGT warming patterns to the hourly sWBGT ERA5 data, we then calculate hourly labor lost in the ‘pattern scaled’ ERA5 data and average the labor lost over this 20-year time period to estimate the mean labor lost in a warmer climate. We assume all parts of the sWBGT diurnal cycle will shift equally. We make this assumption because models do not generally agree on the sign of difference in future changes in the daily maximum vs minimum temperatures, and differences in the magnitude of change in maximum vs minimum are <25%, or <0.2 °C per degree of global warming (Supplementary Fig. 4; Supplementary Text 1).

### Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.