2.1. Employer-employee panel

We source worker-level data on monthly wages from the Annual Social Information Report (RAIS), which is an employer-employee administrative database covering 99 per cent of the formal labor force, starting in January 2015 and ending in December 2016.Footnote ⁹ We draw a random sample of the universe of workers and match them to weather data pertaining to the municipality (county) where they work. We further restrict our sample to workers aged 25 to 55 and exclude workers in agriculture, public administration and military occupations. Our final estimating sample has 3,468,613 worker-month observations representing 209,350 workers.Footnote ¹⁰

We calculate hourly wages as the ratio between monthly wages and monthly hours of work in the main occupation.Footnote ¹¹ To create a measure of real wages (base month-year is January 2015), we deflate nominal wages using the Brazilian official consumer price index by month, calculated by the Brazilian Institute of Geography and Statistics (IBGE).

Although our main focus is on the temperature-wage relationship, we exploit the relationship with other labor market outcomes that could complement our main findings: contractual hours worked per month, days on leave (in a given month) and type of employment contract (permanent versus temporary), as well as changes in firm, sector, and/or municipality of employment.

Table 1 presents descriptive statistics of workers from our estimating sample. The average worker makes, in real terms, about 12.40 Brazilian real (R$) per hour (US$3.10/hour) and has less than a high-school degree (11.44 years of schooling), revealing the relatively low-skill level of the workers in Brazil's formal labor markets.Footnote ¹² From the descriptive statistics we also see that these workers are engaged in long-term contracts since as many as 98 per cent of the sample is employed in permanent jobs. The average worker is employed for an average of 43 hours per week; about 10 per cent of them missed at least one day of work in the 2015–2016 period.

Table 1.

Sample of employees in the formal labor force

Notes: The table presents summary statistics of selected variables calculated for the 209,350 workers in our estimating sample. Real hourly wages are in January 2015 figures. Data source: RAIS 2015–2016.

Table 1 also shows numbers that characterize the panel dimension of the data. Data on wages are available for an average of 16.57 months (out of 24). Wages rose 5.92 times and fell 5.13 times on average over the course of 2015–2016. We also learn that in the two-year period only 3 per cent of workers changed municipalities, 4 per cent changed sector of employment, and 8 per cent switched employers. These low numbers are not surprising given the short time frame our study covers.

Because the vast majority of our sample of formal workers is employed in permanent jobs, it is reasonable to ask how much monthly wages vary within our two-year time span. Figure 1 shows the distribution of the percentage variation in nominal and real wages between consecutive months from January 2015 to December 2016. The panels on the left display all the data, while the panels on the right condition on there being a non-zero change in wages between consecutive months. We observe a spike around zero: 40 per cent of nominal wages do not change and 10 per cent change very little. Excluding zero changes (panels on right), the $10^{th}$ and $90^{th}$ percentiles of the distribution are $-20.7$ and 29.7 per cent respectively. These numbers demonstrate that there is a considerable amount of month-to-month variation in nominal wages. The variation in real wages is even larger because monthly inflation rates are always positive over the period we analyze.

We also note that only a small fraction of the workers in our estimating sample did not experience any changes in nominal wages within the 24-month period (4,612 or 2.2 per cent of the sample). And around 12.8 per cent of the workers (26,860) had their nominal wages change every month within the same period. The deflation of wages naturally creates small variations in real wages even when nominal wages were constant, which explains a larger fraction of workers whose real wages changed every month.Footnote ¹³

2.2. Monthly weather data

To calculate monthly temperature bins, we use daily weather data generated by Xavier et al. (Reference Xavier, King and Scanlon2017).Footnote ¹⁴ The authors use historical weather observations collected from rain gauges and weather stations from the Instituto Nacional de Meteorologia, the Agência Nacional deÁguas, and the Departamento deÁguas e Energia Elétrica de São Paulo. Temperature is available for all Brazilian territory starting in January 1980 and ending in June 2017, and precipitation is available from January 1980 to December 2016. The spatial resolution is 0.25$\times$0.25$^{\circ }$, resulting in 27,216 grids.

We convert grid-level data to municipality-level data as follows. For latitude-longitude pairs which fall inside a municipality's boundaries, we assign as the municipality weather the average weather variables for those grids. When latitude-longitude pairs do not belong to the municipality polygon, we attribute the closest grid to the municipality centroid within a 0.5$^{\circ }$ ($\sim$55 km) range. Using this procedure, we are able to assign weather data to 3,226 municipalities, representing 137 million people, or about 67 per cent of the country's population in 2015.Footnote ¹⁵ After we assign daily grid-level data to each municipality, we distribute daily temperature in eight bin categories, ranging from less than 12$^{\circ }$C to more than 30$^{\circ }$C in intervals of 3$^{\circ }$C. We do the same for precipitation, but using nine bins (see section 3).

Table 2 presents summary statistics of daily temperatures and number of days in a month that fall within each of the eight temperature bins. We calculate these averages from our worker-month sample because that allows us to look at the average number of days in a month that a worker is exposed to a particular temperature range. As the numbers indicate, within a month a worker is exposed to an average of 0.35 days when temperatures are below 12$^{\circ }$C and to 0.40 days when temperatures are above 30$^{\circ }$C. Workers are most exposed to temperatures in the 24–27$^{\circ }$C range – an average of 8.77 days a month.

Table 2.

Average of monthly number of days across eight temperature bins, Brazil and macro-regions

Notes: The table summarizes monthly average daily temperatures and the number of days in eight temperature bins spanning the months from January 2015 to December 2016. Averages are calculated based on the 3,468,613 worker-month observations employed in our estimations. This sample represents 2,436 Brazilian municipalities. Data source: RAIS 2015–2016 and Xavier et al. (Reference Xavier, King and Scanlon2017), 2015–2016.

We also present the averages in table 2 by macro-regions to highlight the extent of the spatial climate variability in Brazil.Footnote ¹⁶ The workers most exposed to lower temperature ranges are located in the South and Southeast; they are the ones contributing to the identification of wage effects in those temperature ranges. Indeed, the average number of days in temperature ranges below 18$^{\circ }$C is zero for those residing in the North and Northeast. Analogously, those most exposed to higher temperatures live in the North and Northeast. For example, workers living in the North experience an average of 22 days a month when temperatures are in the 27–30$^{\circ }$C range, and those in the Northeast, an average of 14 days in that range.

To get a sense of how future changes in climate may alter workers’ exposure to days registering higher temperatures, we simulate the change in the average number of days in the various temperature bins assuming a uniform increase of 2$^{\circ }$C in daily temperatures between 2015 and 2016. We present the simulated averages in figure 2. We see that individuals would experience an average of 2.6 additional days in a typical month when temperatures are above 30$^{\circ }$C and 3.0 more days when temperatures reach levels between 27 and 30$^{\circ }$C. Workers would be exposed to many fewer days with temperatures below 24$^{\circ }$C.

Figure 2.

Average number of days in a month distributed across temperature bins, actual versus uniform climate change scenario.Notes: The figure shows the average number of days in a month for each temperature bin observed in 2015–2016 (light shaded bars) and assuming a flat increase of +2°C across the entire distribution of daily weather to simulate uniform climate change. Averages are calculated from 3,468,613 worker-month observations employed in our estimations. Data source: RAIS 2015–2016 and Xavier et al. (2017), 2015–2016.

3.1. Empirical approach

To estimate the impact of short-term temperature changes on real wages, we adopt an approach based on temperature bins, first employed by Deschenes and Greenstone (Reference Deschenes and Greenstone2011). Because our estimation rests on month-to-month variation in temperature as an identifying source, we ultimately rely on the short-term weather unpredictability for causal inference.

We choose the bins specification below:

(1)\begin{align} \log wage_{ifcmt} & = \beta_0 + \sum_{j = 1}^{8} \beta_j Temp^{\,j}_{cmt} + \sum_{h = 1}^{9} \beta_h Rain^{h}_{cmt}+\alpha_{i} + \alpha_{f} + \alpha_{cm} + \alpha_{ct} + \epsilon_{ifcmt}, \end{align}

where $wage_{ifcmt}$ is the real hourly wage of worker $i$ employed by firm $f$ located in municipality $c$ in month $m$ of year $t$; $Temp^{j}_{cmt}$ is the number of days in municipality $c$, month $m$ of year $t$ that fall in temperature bin $j$. $Rain^{h}_{cmt}$ is the number of days in municipality $c$, month $m$ of year $t$ that fall in precipitation bin $h$. We employ the following temperature bins: below 12$^{\circ }$C; 12 to 15$^{\circ }$C; 15 to 18$^{\circ }$C; 18 to 21$^{\circ }$C (base category); 21 to 24$^{\circ }$C; 24 to 27$^{\circ }$C; 27 to 30$^{\circ }$C; and above 30$^{\circ }$C.Footnote ¹⁷

The remaining parameters, $\alpha _i$, $\alpha _f$, $\alpha _{cm}$ and $\alpha _{ct}$, denote worker, firm, municipality-month, and municipality-year fixed effects, respectively; $\epsilon _{ifcmt}$ is the idiosyncratic error term. These fixed effects are key for identifying the short-term causal effects of temperature on wages. First, municipality-year fixed effects hold constant time-invariant municipality attributes, such as historical and cultural determinants of regional economic development, which are correlated with long-term climate; they also absorb annual shocks to local labor markets that are potentially associated with climate conditions. Municipality-month fixed effects are key to control for weather seasonality that may coincide with economic seasonality that causes labor markets to be more active. Individual fixed effects allow us to account for changes in workforce skill composition, both observed and unobserved, spurred by weather shocks within a municipality. Firm fixed effects help us control for firm-level unobservable characteristics such as endogenous adoption of climate-control technology, as well as changes in the sector/industry mix within a municipality.

3.2. Main results

We present the results in figure 3. Standard errors are clustered by economic region.Footnote ¹⁸ The coefficients on temperature bins lower than the base category (18–21$^{\circ }$C), although negative, are not statistically discernible from the coefficient on the base category. We then conclude that there is no sizable impact on wages from an additional day that falls in those temperature ranges relative to the 18–21$^{\circ }$C range. We do, however, find significant reductions in wages following increases in days that fall in temperature bins above the base category, and the higher the temperatures the higher the point estimates. Notably, the estimates indicate that one day in a month in the 27–30$^{\circ }$C, and not in the 18–21$^{\circ }$C range, leads to a 0.14 per cent reduction in real wages on average; the corresponding impact for the $>$30$^{\circ }$C interval is a 0.2 per cent wage reduction.Footnote ¹⁹ This 0.2 per cent is equivalent to a loss of R$10.44 (US$2.61) in monthly income for the average worker.Footnote ²⁰

Figure 3.

Short-term temperature shocks and real hourly wages.Notes: The figure shows estimates from β coefficients in equation (1) and their 95 per cent confidence interval. We include worker, firm, municipality-month and municipality-year fixed effects, as well as precipitation bins. Temperature bins range from below 12°C to above 30°C in sets of 3°C. The 18–21°C bin is the base category. Standard errors are clustered by economic region. Online appendix table A2, column (1), shows the estimation results. Data sources: labor market data from RAIS 2015–2016 and weather data from Xavier et al. (2017).

We use our estimates to calculate the potential gains in formal sector wages if the daily temperature distribution in 2015 were replaced by that observed in the 1980–2009 period.Footnote ²¹ We estimate a 0.34 per cent gain in average annual wages which, considering the average annual wage of US$6,565.73 in our sample, amounts to nearly US$21.75 per worker. Given the size of the formal sector workforce, this figure adds up to US$1.26 billion in annual wage gains.

Our point estimates are in line with Chen and Yang (Reference Chen and Yang2019), whose analysis of annual firm-level data reveals a 0.21 per cent decline in Chinese industrial output from an additional day at 30$^{\circ }$C relative to the 21–24$^{\circ }$C range. They also compare to the findings reported in Deryugina and Hsiang (Reference Deryugina and Hsiang2014) based on annual county-level data for the U.S. – the authors report a 0.07 per cent reduction in annual net earnings per capita from an additional day above 30$^{\circ }$C relative to the 21–24$^{\circ }$C range. Our results show a 0.12 per cent reduction in hourly wages from an exchange of days between those temperature bins.

Finally, to arrive at a daily-productivity-loss figure, we follow the approach in Deryugina and Hsiang (Reference Deryugina and Hsiang2014).Footnote ²² Assuming that workers’ labor incomes are distributed uniformly across 30 days within a month, each day of work would contribute 1/30 to monthly wages.Footnote ²³ Then, a 0.2 per cent reduction in monthly labor income from an additional day in the $>$30$^{\circ }$C range implies that day is nearly 6 per cent less productive than an average day. Assuming a linear effect above 20$^{\circ }$C – the midpoint of the omitted bin category – yields a marginal effect in daily productivity of $-$0.6%/$^{\circ }$C. This is smaller than the $-$1.7%/$^{\circ }$C effect obtained from annual county-level income data for the U.S. (Deryugina and Hsiang, Reference Deryugina and Hsiang2014), and the $-$1.0%/$^{\circ }$C and $-$2.5%/$^{\circ }$C from annual country-level GDP data for low and middle-income countries (Hsiang, Reference Hsiang2010; Dell et al., Reference Dell, Jones and Olken2012). It is worth noting, however, that these differences are expected since our estimates are obtained from worker-level data and only represent the formal sector of the Brazilian economy.Footnote ²⁴

3.3. Climate change

In this subsection, we use our estimates from figure 3 to perform back-of-the-envelope calculations assuming that daily temperatures have uniformly increased by 2$^{\circ }$C to reflect climate change predictions (Guivarch and Hallegatte, Reference Guivarch and Hallegatte2013; Kwok et al., Reference Kwok, Xu, Liu and Leung2018).

To that end, we compute the predicted change in the number of days in each temperature bin for each municipality (see figure 2) and multiply the change by the estimated coefficient of the corresponding bin. We then impute this change in wages to the observed real wages of each individual in our estimating sample for each month of the year 2015 and sum it across all individuals. Under this flat 2$^{\circ }$C-increase scenario, average annual wages would be 0.87 per cent lower compared to what prevails in 2015.Footnote ²⁵ Extrapolating the losses to the universe of formal workers, we arrive at a R$8.44 billion loss in annual labor income – equivalent to US$2.17 billion. This figure corresponds to 0.12 per cent of 2015 GDP.

To assess regional differences, we consider the same exercise but now separate the sample in two: (i) The North, which encompasses the North and Northeast macro-regions; and (ii) The Center-South, which includes the Midwest, the South and the Southeast macro-regions.Footnote ²⁶ We conclude that the effects are even greater among the warmest and least developed Northern regions: while formal sector workers in the North would lose 1.2 per cent of their monthly wages assuming a uniform increase of 2$^{\circ }$C in temperatures, workers from the Center-South would lose 0.8 per cent.