Articles https://doi.org/10.1038/s41558-019-0632-4 The impact of high ambient temperatures on delivery timing and gestational lengths Alan Barreca   1,2,3* and Jessamyn Schaller3,4 Evidence suggests that heat exposure increases delivery risk for pregnant women. Acceleration of childbirth leads to shorter gestation, which has been linked to later health and cognitive outcomes. However, estimates of the aggregate gestational losses resulting from hot weather are lacking in the literature. Here, we use estimated shifts in daily county birth rates to quantify the gestational losses associated with heat in the United States from 1969 to 1988. We find that extreme heat causes an increase in deliveries on the day of exposure and on the following day and show that the additional births were accelerated by up to two weeks. We estimate that an average of 25,000 infants per year were born earlier as a result of heat exposure, with a total loss of more than 150,000 gestational days annually. Absent adaptation, climate projections suggest additional losses of 250,000 days of gestation per year by the end of the century. I ncreased exposure to hot weather with climate change is likely to harm infant health1, although the magnitude of the threat has not been well documented for the United States as a whole. Infant health is multifaceted, but earlier delivery is one important metric that has been shown to be strongly associated with neonatal health and cognitive outcomes later in childhood2,3. While extreme heat might have a latent effect on gestational length through exposure early in pregnancy4, the focus of our study is on the contemporaneous relationship between ambient temperature and delivery risk towards the end of pregnancy. Extreme heat has been hypothesized to cause an immediate increase in delivery risk, both in animals and in humans5–8. One possible channel for these effects is that heat increases levels of oxytocin6,7, which is a key hormone regulating the onset of delivery. Alternatively, extreme heat might cause earlier deliveries via cardiovascular stress8. Here, we estimate the aggregate loss of gestational days due to hot weather across the United States during a contemporary period and then use those results to forecast the impacts of climate change on gestational lengths. Our 20 yr sample (1969–1988) encompasses 56 million births spanning more than 3 million county days, which is far greater spatial and temporal coverage than other studies correlating temperature with gestational lengths9–18. Existing studies have examined the association between temperature and contemporaneous gestational lengths as recorded on birth certificates9–18. These studies have generally found an association between hot weather and shorter gestational lengths (see refs. 19,20 for an overview). However, they do not account for compositional changes in births due to shifts in delivery dates across the exposure window. In studies exploring impacts on gestational length, hot weather might cause later-term pregnancies (>41 weeks) to shift forwards by a few days, which could increase observed gestational lengths in the exposure window13,15. In studies focusing on risk of pre-term delivery (<37 weeks), an increase in pre-term delivery risk in the exposure window may reflect a shift among births that would have qualified as pre-term even had they been born after the exposure window9–12,14,16–18. One study using monthly US data finds suggestive evidence of a reduction in gestational length, but the coarse temporal nature of the data precludes calculation of gestational losses21. In summary, the existing studies cannot provide accurate estimates of gestational losses, although they do provide evidence of the existence of a relationship. Given recent increases in the frequency of extremely hot weather, there is a clear need to better forecast the potential magnitude of climate change’s impact on infant health at the national level22. We advance the methodology of temperature–gestation studies by using data on daily birth rates, as opposed to recorded gestational lengths. Our approach can quantify the loss of gestational days that results from heat exposure by tracing out the shift in timing of births across days. For example, in a given county of the United States, an increase in birth rates on the day of hot weather followed by a decrease two days later suggests temperature reduced gestational lengths by two days. In addition to providing a more accurate estimate of gestational losses than previous work, using shifts in daily births is an important innovation for two reasons. First, our method allows for estimation of the effects of heat on delivery timing in nations that lack data on gestational lengths but do record dates of birth. Second, when present, recorded gestational lengths are likely to suffer from misreporting, which is often linked to low socioeconomic status in a way that might lead to underestimation of population-wide impacts23. Results The main outcome variable in our study is the daily birth rate, defined as number of births in a given county per 100,000 women aged 15–44 yr, which averages 19.1 in our sample (Table 1). We present the average frequencies of temperatures in each of seven temperature ‘bins’, defined by maximum temperatures <40, [40–50), [50–60), [60–70), [70–80), [80–90) and ≥90 °F (Table 1). These bins correspond approximately to <4.4, [4.4–10.0), [10.0–15.6), [15.6–21.1), [21.1–26.7), [26.7–32.2) and ≥32.2 °C. In our full 1969–1988 sample, 8.3% of days, or about 30.3 days per year, have a maximum temperature ≥90 °F (32.2 °C). We also average across four geographic regions. Exposure to extreme heat and birth rates varies regionally: counties in the South and West Census Regions Institute of the Environment and Sustainability, University of California–Los Angeles, Los Angeles, CA, USA. 2IZA Institute of Labor Economics, Bonn, Germany. 3National Bureau of Economic Research, Cambridge, MA, USA. 4Robert Day School of Economics and Finance, Claremont McKenna College, Claremont, CA, USA. *e-mail: abarreca@ioes.ucla.edu 1 Nature Climate Change www.nature.com/natureclimatechange Articles NATUrE ClimATE CHAngE Table 1 Summary statistics on daily birth rates and daily weather for 1969–1988 Sample All states Northeast Midwest South West Birth rate 19.092 17.047 19.288 19.699 20.204 Maximum temperature (°F) <40 0.128 0.190 0.234 0.044 0.050 Maximum temperature (°F) [40–50) 0.087 0.128 0.103 0.060 0.062 Maximum temperature (°F) [50–60) 0.120 0.147 0.114 0.099 0.133 Maximum temperature (°F) [60–70) 0.181 0.173 0.151 0.163 0.264 Maximum temperature (°F) [70–80) 0.175 0.164 0.150 0.185 0.205 Maximum temperature (°F) [80–90) 0.228 0.173 0.202 0.311 0.181 Maximum temperature (°F) ≥90 0.083 0.024 0.046 0.139 0.106 No precipitation 0.707 0.653 0.679 0.710 0.809 Precipitation (0.0–0.5] inches 0.225 0.266 0.260 0.207 0.156 Precipitation  >0.5+ inches 0.068 0.081 0.062 0.084 0.035 Raw number of observations (births) 56,030,127 11,079,223 15,445,315 19,726,346 9,779,243 Number of county days 3,250,725 759,720 854,685 1,081,140 555,180 Note: For temperature and precipitation, the data represent the fraction of the year in a given range. Estimates are weighted by the average county population. The birth rates are per 100,000 women aged 15–44 yr. We exclude Alaska and Hawaii from our sample to focus on the continental United States. Northeast, Midwest, South, and West are the four Census Regions of the United States. experience more hot days and have higher birth rates (Table 1). The cross-region correlation between temperature and birth rates that we observe is probably driven in part by socioeconomic factors that are spuriously correlated with temperature. We address this in our empirical model. To isolate the causal effects of temperature on the timing of childbirth, we focus on plausibly unpredictable variation in temperature exposure for a particular county and time of year. Our model includes county by day-of-year fixed effects, which allow us to disentangle the causal effects of temperature from regional differences in demographics, socioeconomic status or health behaviours (such as smoking prevalence) and from any seasonal confounders. Intuitively, our model compares birth rates on a particular hot day in a given county with birth rates on the same calendar day in the same county in other years when that day was not as hot, while controlling for nationwide changes in birth rates over time. In addition to reflecting the direct effects of ambient heat, our temperature-exposure variables incorporate indirect causal channels through which temperature might impact delivery risk, such as temperature-driven changes in pollution, such as ozone, or behaviour, such as spending time outdoors. We allow for a nonlinear relationship between temperature and delivery risk by including the seven temperature-exposure categories listed in Table 1, with [60–70) °F (or [15.6–21.1) °C) as the reference category. The model includes 30 days of lagged temperature bins so we can quantify the shifts in daily birth rates—a key innovation of our work. The spirit of our empirical design builds on recent research on the temperature–fertility relationship21 as well as other research in the climate-economics literature24. More details on the model can be found in the figure notes and Methods. To aid in the exposition of the results, we first show the contemporaneous effects of temperature (each of the seven maximum temperature bins, controlling for temperature bins on the 30 preceding days) on the birth rate. We find that birth rates increase by 0.97 births per 100,000 women aged 15–44 yr on days with a maximum temperature ≥90 °F (32.2 °C) (Fig. 1a), compared with days with a maximum [60–70) °F temperature. This reflects an increase of 5% relative to the average daily county birth rate. Birth rates are also elevated on days with a maximum temperature [80–90) °F (or [26.7–32.2) °C), although only by 0.57. In general, we find that the association between daily maximum temperature and contemporaneous county birth rates is reverse J-shaped, mostly flat up to [60–70) °F (or [15.6–21.1) °C) with a kink upwards thereafter. We explore the sensitivity of these results to how temperature exposure is modelled by using a spline in the daily maximum temperature (Fig. 1b) and adding a ≥100 °F (37.8 °C) temperature category (Fig. 1c). These results show that the contemporaneous relationship between temperature and birth rates is similar regardless of how temperature is modelled and that the effect is increasing past 90 °F. The qualitative conclusions are similar if we use daily minimum temperature in place of daily maximum temperature (Fig. 1d). We also test for impacts using daily mean temperature (Supplementary Fig. 1). In results not reported, we estimate a model with both daily maximum and daily minimum temperatures and find that, in fact, the daily minimum temperatures are more predictive using an F test of joint significance. This suggests that nighttime exposure, and possibly disrupted sleep25, could be more important than daytime exposure. Nonetheless, we opt to use the daily maximum temperature in our core model due to its ease of interpretation and popularity in public discourse surrounding the weather. We continue with the same empirical model as that in Fig. 1a, but now show the estimated effects of exposure to a ‘hot day’—a day with a maximum temperature ≥90 °F (32.2 °C)—on county birth rates on the day of the temperature shock and on each of the following 30 days. We continue to control for the other temperatureexposure categories over the full set of days, so the estimates should still be interpreted as the difference between a ≥90 °F (32.2 °C) day and a [60–70) °F (or [15.6–21.1) °C) day. The results show that county birth rates are 0.97 higher on a hot day compared with a [60–70) °F (or [15.6–21.1) °C) day (Fig. 2a), which is identical to the estimate in Fig. 1a by design. Birth rates are also elevated by 0.66 on the day following a hot day (+1 day from exposure), which could reflect a delayed effect or could be due in part to the fact that labour and delivery can carry over from one day to the next. Two days after exposure, birth rates decrease by 0.57, which suggests that a sizable fraction of affected births experience a shift in delivery date of as little as one or two days. The coefficients on days 3 to 15 are negative, although most of the coefficients have 95% confidence intervals that include zero. These decreases represent an absence of births, evidence of a shift in delivery date. The cumulative effect of exposure to extreme temperature on county birth rates is calculated for each day following heat exposure by summing the coefficients on the contemporaneous and lagged temperature variables up to that day (Fig. 2b). Intuitively, for a temperature shock occurring on day t, the cumulative effect j days later Nature Climate Change www.nature.com/natureclimatechange Articles NATUrE ClimATE CHAngE a b 1.5 1.5 1.0 Birth rate Birth rate 1.0 0.5 0 0.5 0 –0.5 –0.5 <40 40–50 50–60 60–70 70–80 80–90 ≥90 40 50 c d 1.5 70 80 90 1.5 1.0 Birth rate 1.0 Birth rate 60 Temperature (°F) Temperature (°F) 0.5 0.5 0 0 –0.5 –0.5 <40 40–50 50–60 60–70 70–80 80–90 90–100 ≥100 Temperature (°F) <20 20–30 30–40 40–50 50–60 60–70 ≥70 Temperature (°F) Fig. 1 Temperature–dose response function on day 0. Each panel represents estimates generated from a separate model. Other than how temperature exposure is constructed, the outcome and control variables are the same in all four panels. The outcome is the daily birth rate per 100,000 women aged 15–44 yr, which has a mean of 19.1 in our sample. The model has year-day fixed effects, county by day-of-year fixed effects, county by month-of-year specific linear time trends and state-by-year-month fixed effects. We control for whether the day’s precipitation is (0.0-0.5] inches or > 0.50 inches. Estimates are weighted by average county population. Standard errors are clustered at the state level. Only day 0 is plotted here, but the models control for the temperature exposure on days −5 through 30, as illustrated in Fig. 2. a, Our core model includes daily maximum temperature bins. b, A cubic polynomial spline with knots at 20, 40, 60, 80 and 100 °F of daily maximum temperature; the reference temperature is 65 °F. c, Our core model, but the maximum bin is set at 100 °F. d, Daily minimum temperature is used in place of maximum temperature to define the bins. The error bars represent ±2 standard errors. reflects the change in the county birth rate for the entire period from t to t + j that resulted from that temperature shock. The cumulative effect is largest on the first day after the shock (day 1), reflecting an increase in the birth rate of 1.63 births. Starting on day 2, the cumulative effect of a hot day begins to fall. The 95% confidence interval includes zero on day 13 and the estimate is nearly equal to zero on day 15. This suggests that the majority of intertemporal shifts in deliveries are complete within two weeks, although we cannot rule out that there may also be larger gestational losses for a small number of births. We use the estimated cumulative effects (Fig. 2b) to conduct calculations that provide a sense of the aggregate number of days of lost gestation following heat exposure. First, we calculate that one hot day (≥90 °F or ≥32.2 °C) results in 9.9 lost gestational days per 100,000 women aged 15–44 yr in a county to day 15. Multiplying this by an approximation of the average population of women aged 15–44 yr during our sample period (approximately 50.45 million), we estimate that just one additional hot day per year, if experienced nationwide, would result in approximately 5,000 total lost days of gestation. To generate the total annual loss of gestational days resulting from heat exposure, we multiply this estimate by the average number of hot days per year (30.3 days), arriving at approximately Nature Climate Change www.nature.com/natureclimatechange 151,000 lost days of gestation on average per year in the United States over our sample period. Given that we observe only county-level birth rates and not flows of individual births, it is impossible to generate the exact number of births that actually occurred sooner or later than they otherwise would have as a result of heat exposure. However, by treating the combined increase in births on day 0 and day 1 as a lower-bound estimate of the number of forward-displaced births, we can generate a conservative upper-bound estimate of the average loss of gestation per affected infant. First, we divide the total number of lost gestational days by this combined increase (1.63 births per 100,000 women aged 15–44 yr). Then, multiplying this by the average population of women aged 15–44 yr, we find that one additional hot day per year affects at least 822 births. Multiplying again by the average number of hot days per year, we estimate that, at a minimum, just under 25,000 births per year occurred earlier than otherwise during our sample period. Finally, we combine this estimate with our previous estimates of the number of lost days of gestation and estimate 6.1 lost gestational days per infant affected by the ≥90 °F (32.2 °C) day. Considering the cumulative effects estimated separately for white mothers and black mothers shows that there is a greater impact on Articles NATUrE ClimATE CHAngE a 1.0 Birth rate 0.5 0 –0.5 –1.0 –5 0 5 10 15 20 25 30 20 25 30 Days from exposure b 2.0 Birth rate 1.5 1.0 0.5 0 –0.5 –5 0 5 10 15 Days from exposure Fig. 2 Effect of daily maximum temperature ≥90 °F on birth rates by days from exposure. The model estimates come from the same model as Fig. 1. The model has seven temperature-exposure bins. However, here only the estimated effect of maximum temperature ≥90 °F is presented. a, Distributed lag coefficients. Error bars represent ±2 standard errors, with standard errors clustered at the state level. b, Cumulative effects. The solid (dashed) line represents the estimated cumulative effect (±2 standard errors) on the birth rate between day zero and a given day. The grey shading highlights the contemporaneous effect. The estimates can be interpreted as the impact on the daily birth rate from the day having a maximum temperature ≥90 °F, relative to [60–70) °F. The negative days represent placebo checks since the births occurred before temperature exposure. We exclude the placebo estimates in b. black mothers (Fig. 3a). The effect on day 0 is similar across the two groups, but the cumulative effect is larger after day 1 for black mothers. Furthermore, the rebound in births occurs through day 30 for black mothers as opposed to day 15 for white mothers, implying that the loss in gestational length is greater for black mothers. We test for the potential role of adaptation in mitigating the effects of temperature exposure on the timing of births in two ways. First, we explore whether the effects of temperature on birth rates are dampened in places where hot weather is more common. People living in hotter climates could have acclimatized through natural physiological changes, adoption of cooling technologies or changes in building designs and urban planning, among other things. We stratify counties into two groups on the basis of the frequency of days with maximum temperatures ≥90 °F (32.2 °C) during our study period. The cumulative effects of heat on birth timing are stronger in counties with lower frequency of hot-weather exposure (cold counties) (Fig. 3b). In counties with more-frequent hot-weather exposure (hot counties), the maximum cumulative effect is lower and returns to zero faster after the shock. It is worth noting that geographic differences in the vulnerability to heat might also incorporate differences in wealth or other confounders. Nonetheless, the pattern of effects is consistent with adaptation. Second, we test for the importance of air conditioning in mitigating the association between temperature exposure and birth outcomes following previous studies21,26,27. During our sample period (1969–1988), air conditioning was adopted rapidly: the fraction of the population with residential air conditioning increased from 37% in 1970 to 55% in 198028. Adoption of air conditioning is also likely related to other factors that might influence prenatal health and resiliency to hot weather, so we view these estimates as suggestive of air conditioning’s causal impact. When we include a set of interaction terms between daily maximum temperature bins and county-level residential air conditioning coverage in our model, we find that the interaction term for ≥90 °F (32.2 °C) days is imprecisely estimated (Fig. 4a). As residential air conditioning may be important for mitigating nighttime heat exposure, we also interact air conditioning with daily minimum temperatures (Fig. 4c). The interaction term for ≥70 °F (21.1 °C) days in the model with minimum temperature is negative and quite large: full air conditioning coverage mitigates three-quarters of the impact of hot days as represented by minimum temperature. This may be because nighttime exposure poses a greater threat to prenatal health or because the affected population more effectively utilizes air conditioning at nighttime. Finally, we assess the implications of our estimates for climate change. We take output from 22 modern climate models (Climate Model Intercomparison Project phase 5 (CMIP5))29 and calculate the within-model changes in the frequency of ≥90 °F days between the 2000–2019 and 2080–2099 periods. We then project our estimates using the average changes across the 22 models. At the end of the century (2080–2099), we estimate that there will be approximately 253,000 additional lost days of gestation per year on average in the United States, affecting nearly 42,000 additional births (see Supplementary Fig. 3 for year-by-year projections). If we account for the full distribution of temperature changes, the gestational losses increase to 316,000 days. These projections hold adaptive investments, such as air conditioning, constant, so these should be considered upper-bound estimates of the gestational losses. Discussion Using a large dataset from the United States, we find that exposure to extreme heat causes a large increase in delivery risk. We estimate that birth rates increase by 5% on days with maximum temperatures above 90 °F (32.2 °C), which is similar to a result from one recent study in Montreal10. At least 25,000 births per year experience some reduction in gestational length in the United States over our sample period. The average reduction in gestational length is modest (6.1 days), although some births experience a loss of two weeks. Direct comparisons of our gestational loss estimates with past research is difficult given that previous estimates fail to account for compositional changes due to temperature-induced shifts in birthdates. For example, one study found that average gestational length fell by five days due to exposure to extreme heat on the previous day13, but it is possible that such a large decrease in the average gestational length was driven by a one- to two-day acceleration of births that would have still had lower-than-average gestational lengths in the absence of extreme heat exposure. While past work has demonstrated the existence of a relationship between heat and birth timing, our study provides a reliable estimate of the causal effect of temperature on gestational lengths using nationally representative data from the United States. Nature Climate Change www.nature.com/natureclimatechange Articles NATUrE ClimATE CHAngE 1.5 1.5 1.0 Black White Birth rate b 2.0 Birth rate a 2.0 1.0 Cold counties Hot counties 0.5 0.5 0 0 –5 0 5 10 15 20 25 30 –5 0 5 Days from exposure 10 15 20 25 30 Days from exposure Fig. 3 Heterogeneous effects of daily maximum temperature ≥90 °F on birth rates, cumulative effects by days from exposure. a, Black (solid line) versus white (dashed line) mothers. b, Cold (solid line) versus hot (dashed line) counties. ‘Cold counties’ are below the median number of days with daily maximum temperature ≥90 °F over our sample period, while ‘hot counties’ are above and including the median. The model is similar to Fig. 2, except we estimate the model separately for black and white mothers (a) and for two groups of counties on the basis of their climate (b). We drop counties with fewer than 100,000 population from the analysis in b. Standard errors are excluded from this figure to facilitate presentation. a b 0.5 1.5 Maximum temperature main effect 1.0 Birth rate Birth rate 0 0.5 Maximum temperature interaction with air conditioning –0.5 0 –0.5 –1.0 <40 40–50 50–60 60–70 70–80 80–90 ≥90 <40 40–50 c d 0.5 70–80 80–90 ≥90 60–70 ≥70 Minimum temperature main effect 2 –0.5 Birth rate Birth rate 60–70 3 0 –1.0 Minimum temperature interaction with air conditioning –1.5 50–60 Temperature (°F) Temperature (°F) 1 0 –2.0 –1 <20 20–30 30–40 40–50 50–60 60–70 ≥70 Temperature (°F) <20 20–30 30–40 40–50 50–60 Temperature (°F) Fig. 4 Effect of daily maximum temperature ≥90 °F relative to 60–70 °F on birth rates, interacted with air conditioning. a,b, Model with maximum temperature, where we allow for an interaction between maximum temperature and residential air conditioning coverage (a) as well as including the main effect of maximum temperature (b). c,d, Model with minimum temperature. c and d are similar to a and b, respectively, but use minimum temperature in place of maximum temperature. The air conditioning variable is linearly interpolated between the 1960, 1970 and 1980 Decennial Censuses and represents the fraction of households with an air conditioning unit. To improve precision, these models include temperature exposure on days 0 to 15 (as opposed to −5 to 30 as with our core model). However, we show only the effect at day 0 here. The error bars represent ±2 standard errors, with standard errors clustered at the state level. These findings are important given the emerging evidence that early-life health has a lasting effect on health and cognitive outcomes4. Using administrative data on adult earnings in the United States, one recent study27 found that exposure to extremely hot Nature Climate Change www.nature.com/natureclimatechange days during the third trimester (among other points in the prenatal and postnatal periods) leads to lower income in adulthood. Our study suggests that shortened gestational lengths might be one mechanism underlying this relationship. However, the effects of Articles temperature are not likely confined to earlier delivery risk. There is existing evidence that exposure to extreme heat in the second and third trimesters leads to lower birth weights even after controlling for gestational length30. While we posit climate change will cause gestational losses, the exact magnitude of the future costs is highly uncertain—households may adapt as expectations about the frequency of hot weather events increase, which could mitigate impacts on infant health. Indeed, we find that access to air conditioning is an effective adaptation strategy and one that is likely to be adopted more in places where hot weather is currently infrequent. But given air conditioning use would increase GHG emissions, more research is needed to evaluate whether there exist alternative adaptation strategies for pregnant women. Online content Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41558019-0632-4. Received: 3 June 2019; Accepted: 18 October 2019; Published: xx xx xxxx References 1. Poursafa, P., Keikha, M. & Kelishadi, R. Systematic review on adverse birth outcomes of climate change. J. Res. Med. Sci. 20, 397–402 (2015). 2. Figlio, D. N., Guryan, J., Karbownik, K. & Roth, J. Long-term cognitive and health outcomes of school-aged children who were born late-term vs full-term. JAMA Pediatr. 170, 758–764 (2016). 3. Spong, C. Y. Defining “term” pregnancy: recommendations from the Defining “Term” Pregnancy Workgroup. JAMA 309, 2445–2446 (2013). 4. Almond, D. & Currie, J. Killing me softly: the fetal origins hypothesis. J. Econ. Perspect. 25, 153–172 (2011). 5. Tao, S., Monteiro, A. P. A., Thompson, I. M., Hayen, M. J. & Dahl, G. E. Effect of late-gestation maternal heat stress on growth and immune function of dairy calves. J. Dairy Sci. 95, 7128–7136 (2012). 6. Dreiling, C. E., Carman, F. S. & Brown, D. E. Maternal endocrine and fetal metabolic responses to heat stress. J. Dairy Sci. 74, 312–327 (1991). 7. Fuchs, A., Fuchs, F., Husslein, P., Soloff, M. S. & Fernstrom, M. J. Oxytocin receptors and human parturition: a dual role for oxytocin in the initiation of labor. Science 215, 1396–1398 (1982). 8. Hampel, R. et al. Short-term impact of ambient air pollution and air temperature on blood pressure among pregnant women. Epidemiology 22, 671–679 (2011). 9. Strand, L. B., Barnett, A. G. & Tong, S. Maternal exposure to ambient temperature and the risks of preterm birth and stillbirth in Brisbane, Australia. Am. J. Epidemiol. 175, 99–107 (2011). NATUrE ClimATE CHAngE 10. Auger, N., Naimi, A. I., Smargiassi, A., Log, E. & Kosatsky, T. Extreme heat and risk of early delivery among preterm and term pregnancies. Epidemiology 25, 344–350 (2014). 11. Avalos, L. A., Chen, H., Li, D. & Basu, R. The impact of high apparent temperature on spontaneous preterm delivery: a case-crossover study. Environ. Health 16, 5 (2017). 12. Basu, R., Malig, B. & Ostro, B. High ambient temperature and the risk of preterm delivery. Am. J. Epidemiol. 172, 1108–1117 (2010). 13. Dadvand, P. et al. Climate extremes and the length of gestation. Environ. Health Perspect. 119, 1449–14453 (2011). 14. Lee, S. J., Hajat, S., Steer, P. J. & Filippi, V. A time-series analysis of any short-term effects of meteorological and air pollution factors on preterm births in London, UK. Environ. Res. 106, 185–194 (2008). 15. Porter, K. R., Thomas, S. D. & Whitman, S. The relation of gestation length to short-term heat stress. Am. J. Public Health 89, 1090–1092 (1999). 16. Schifano, P. et al. Effect of ambient temperature and air pollutants on the risk of preterm birth, Rome 2001–2010. Environ. Int. 61, 77–87 (2013). 17. Vicedo-Cabrera, A. M., Iniguez, C., Barona, C. & Ballester, F. Exposure to elevated temperatures and risk of preterm birth in Valencia, Spain. Environ. Res. 134, 210–217 (2014). 18. Yackerson, N., Piura, B. & Sheiner, E. The influence of meteorological factors on the emergence of preterm delivery and preterm premature rupture of membrane. J. Perinatol. 28, 707–711 (2008). 19. Strand, L. B., Barnett, A. G. & Tong, S. The influence of season and ambient temperature on birth outcomes: a review of the epidemiological literature. Environ. Res. 111, 451–462 (2011). 20. Song, X. et al. Impact of ambient temperature on morbidity and mortality: an overview of reviews. Sci. Total Environ. 586, 241–254 (2017). 21. Barreca, A., Deschenes, O. & Guldi, M. Maybe next month? Temperature shocks and dynamic adjustments in birth rates. Demography 55, 1269–1293 (2018). 22. IPCC Climate Change 2014: Synthesis Report (eds Core Writing Team, Pachauri, R. K. & Meyer, L. A.) (IPCC, 2014). 23. Lynch, C. D. & Zhang, J. The research implications of the selection of a gestational age estimation method. Paediatr. Perinat. Epidemiol. 21, 86–96 (2007). 24. Dell, M., Jones, B. F. & Olken, B. A. What do we learn from the weather? The new climate–economy literature. J. Econ. Lit. 52, 740–798 (2014). 25. Obradovich, N., Migliorini, R., Mednick, S. C. & Fowler, J. H. Nighttime temperature and human sleep loss in a changing climate. Sci. Adv. 3, e1601555 (2017). 26. Barreca, A., Clay, K., Deschenes, O., Greenstone, M. & Shapiro, J. S. Adapting to climate change: the remarkable decline in the US temperature–mortality relationship over the twentieth century. J. Polit. Econ. 124, 105–159 (2016). 27. Isen, A., Rossin-Slater, M. & Walker, R. Relationship between season of birth, temperature exposure, and later life wellbeing. Proc. Natl Acad. Sci. USA 114, 13447–13452 (2017). 28. Biddle, J. Explaining the spread of residential air conditioning, 1955–1980. Explor. Econ. Hist. 45, 402–423 (2008). 29. Taylor, K. E., Stouffer, R. J. & Meehl, G. A. An overview of CMIP5 and the experiment design. Bull. Am. Meteorol. Soc. 93, 485–498 (2012). 30. Deschênes, O., Greenstone, M. & Guryan, J. Climate change and birth weight. Am. Econ. Rev. 99, 211–217 (2009). Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. © The Author(s), under exclusive licence to Springer Nature Limited 2019 Nature Climate Change www.nature.com/natureclimatechange Articles NATUrE ClimATE CHAngE Methods Data. The primary data for our analysis span the 1969–1988 period and come from two sources: natality data from National Vital Statistics (NVS) and weather data from Global Historical Climatology Network (GHCN). We focus on the period between 1969 and 1988 since date of birth is only available in those years in the public NVS. The main outcome is the daily county birth rate. We construct birth rates by dividing the number of births on a given county day by the estimated population of women aged 15 to 44 yr in that county year. The population data come from National Cancer Institute’s Surveillance, Epidemiology, and End Results (SEER) programme. Total births come from the NVS, which has information on county and date of birth between 1969 and 1988. The NVS data are 50% samples 1969 to 1971 for all states, a mixture of 50% and 100% samples across states between 1972 and 1984 and 100% samples for all states and years after 1985. County is assigned on the basis of the reported residence of the mother. We match daily weather records to the birth records by exact date of birth and county of mother’s residence. The weather data are constructed from the GHCN daily (GHCND) data, which are maintained by the National Oceanic and Atmospheric Administration. The GHCND data have information on (daily) maximum and minimum temperature and total precipitation for about 4,000 weather stations throughout the United States during our time period. The key temperature exposure variables are one of seven temperature ‘bins’, defined as stations where the day’s maximum temperature is below 40, 40–50, 50–60, 60–70, 70–80, 80–90 or above 90 °F. These bins correspond to below 4.4, 4.4–10, 10.0–15.6, 15.6–21.1, 21.1–26.7, 26.7–32.2 and above 32.2 °C. We calculate weather conditions at the county level using weather stations within 100 miles of the county centroid and inverse distance weights. While we focus on maximum temperature, much of the existing literature relies on mean temperature. For this reason, we present results using daily mean temperature bins in Supplementary Fig. 1. Data on humidity, which we use in Supplementary Fig. 2, come from a separate source (Global Surface Summary of the Day) and have more limited coverage than the temperature data. County-level air conditioning data come from the 1960, 1970 and 1980 decennial censuses. Following a recent study26, we linearly interpolate air conditioning coverage between census years and extrapolate out to 1988, the last year of our sample. We group counties with fewer than 100,000 people (on average) in each state for computing efficiency. We also group several counties to be consistent with the population identifiers in the SEER data, which mainly affects the counties on Manhattan Island in New York. Note that this aggregation should only affect precision since we model the nonlinear effects of temperature at the station level before aggregation. This reduces the number of individual counties to 397 with another 48 county groups for a total of 445 geographic units. In these instances, we assign weather exposure at the county level and then aggregate to the county group using the county populations as weights, whereas we simply sum the total births and divide by total population of women aged 15–44 yr in the county group for the birth-rate variable. Assuming the empirical model is properly specified at the county level, such aggregation should affect only the precision of our estimates. Note that approximately 70% of births in the United States reside in counties with over 100,000 population. The results are similar when we drop the less-populated counties altogether. For our climate change analysis, we took output from 22 climate models from CMIP529 and calculated the average within-model changes in the distribution of temperatures over the twenty-first century. The 22 models are ACCESS1–0, ACCESS1–3, CMCC-CESM, CMCC-CM, CMCC-CMS, CNRM-CM5, MIROCESM-CHEM, MPI-ESM-LR, MPI-ESM-MR, MRI-CGCM3, CanESM2, CCSM4, CESM1-BGC, CESM1- CAM5, CSIRO-Mk3–6–0, GFDL-CM3, GFDL-ESM2G, GFDL-ESM2M, IPSL-CM5A-LR, IPSL-CM5A-MR, IPSL- CM5B-LR and MIROC5. The models were representative concentration pathway (RCP) 8.5 or ‘business as usual’. We calculate national averages using county population weights, where we take the nearest grid point to each county in the United States. We drop Alaska and Hawaii to focus on the continental United States. We then project our estimates using the average changes in the temperature distribution across these 22 models. We calculate the changes using the internal changes in the model’s own output using the 2000–2019 period as the baseline. Estimation strategy. Following the general approach laid out in one recent study21, we estimate the effects of temperature on delivery risk with the following model via ordinary least squares: Yct ¼ ΣL βL f ðTMAXÞct�L þPRCPct þ DATEt þ δcd þ αcm ´ YEAR þ Ssym þ ect ð1Þ where Y is the number of births per 100,000 women between 15 and 44 yr on date t in county c; β is the main set of parameters to be estimated that are allowed to vary over the sum Σ days L; f(TMAX) flexibly models the maximum temperature on date t – L to allow temperature to have a contemporaneous effect on the outcome as well as a lagged effect for up to 30 days (L = 0, 1, 2, …, 30); we also include five leads of temperature as a placebo check (L = −1, −2, …, −5). PRCP controls for days with rainfall (0-0.5] inches and rainfall >0.5 inches, with no rainfall as the omitted category, to account for the possibility that rainfall is correlated with Nature Climate Change www.nature.com/natureclimatechange both temperature and birth rates; DATE is a set of year-day fixed effects to help improve precision and avoid spurious time-series correlation between births and temperatures; note that DATE implicitly controls for weekday/weekend and holiday effects in a flexible way that is allowed to vary over time. County by day-of-year (d) fixed effects δ (for example, 8 August in Los Angeles County) mitigate potential biases from county-specific seasonal changes in fertility patterns that might be correlated with expected climatic conditions; α is a set of county by month-of-year (m) specific time trends (for example, Augusts in Los Angeles County interacted with linear year variable) to address trends in birth seasonality over time that may be spuriously correlated with gradual climatic trends at the county level. We also control for state-year-month fixed effects (Ssym) to mitigate noise from state-level policy changes, with the s subscript refering to state, y the calendar year, and m the calendar month. We weight the regressions by the average population of women aged 15–44 yr in each county over the sample period and cluster the standard errors at the state level to allow the error term (e) to be spatially and temporally correlated within states. In our core model, we rely on a ‘binned’ approach to modelling the temperature-response function that is common in the literature24. We categorize a station day by whether the daily maximum temperature falls in a given range, that is, <40, [40–50), [50–60), [60–70), [70–80), [80–90), ≥90 °F. These bins correspond to <4.4, [4.4–10.0), [10.0–15.6), [15.6–21.1), [21.1–26.7), [26.7–32.2) and ≥32.2 °C. The binned approach has the advantage of both being easier to interpret and imposing relatively few functional form assumptions. We have tested robustness of our results to modelling temperature in other ways, including using a spline, using the minimum or mean temperature when constructing the bins and controlling for both temperature and humidity. To test the validity of our identification strategy, we also examine coefficients on temperature in the five days after the realized births (the negative days in Fig. 1a). While it is possible that anticipatory responses to weather forecasts might alter maternal behaviour, future hot temperatures cannot generate a direct biological response. Thus, an association between future temperatures and birth rates today is unlikely to reflect a causal biological effect and may generate concerns about the specification. We find no meaningful relationship between future temperatures and contemporaneous birth rates, although one of the five coefficients (day −5) has a 95% confidence interval that excludes zero. Our hypothesis is that hot weather causes intertemporal shifts in delivery dates, but in potentially heterogeneous ways across subpopulations. For example, hot weather may cause some deliveries to shift forwards one day, from tomorrow (t = 1) to today (t = 0), while other deliveries might shift forwards by two days, from the day after tomorrow (t = 2) to today (t = 0). There could also be latent effects that cause some births to shift forwards at some point in the future, for example from the ten days (t = 10) later to five days later (t = 5). Given that we do not observe the true counterfactual day of birth, the precise shift in gestational length from the temperature shock for each subpopulation is not observable. However, we can estimate the loss of gestational days across the whole population associated with a temperature shock by taking the cumulative effect of the β coefficients in equation (1). Specifically, the total loss of gestational days in a given county is given by the formula ΣDd=1 d × (Δd) where d is the number of days since the shock and Δd is the change in the number of births occurring d days after the shock. If we divide this formula by the average population for that county, the result will be the summation of the β coefficients from equation (1). For this calculation, we set D to 30 days and show that the shifts are mostly confined to 15 days from the temperature shock. We also use this estimate of the total loss of gestational days to estimate an upper bound on the average loss of gestational length per impacted birth, dividing it by the sum of the day 0 and day 1 coefficients. In Supplementary Fig. 2, we show estimates that control for both temperature and specific humidity, which measures the amount of water vapour in the air. We find that hot and humid weather causes more early deliveries than hot and dry days, although the qualitative results for temperature are similar, with a tipping point near 60–70 °F (15.6–21.1 °C). We opt for the model that controls only for temperature to aid in the exposition of the results. Reporting Summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability All data necessary for replication of the results in this paper are available for download at https://figshare.com/s/c984d29316e012b79e32. The original weather data can be downloaded from ftp://ftp.ncdc.noaa.gov/pub/data/ghcn/daily. The original birth records can be downloaded from http://www.nber.org/natality/. The population data can be downloaded from https://seer.cancer.gov/popdata/ download.html. Interested researchers should contact Urs Beyerle or Jan Sedlacek at ETH Zurich to gain access to the CMIP5 climate projection data. Source data for Figs. 1–4 are provided with the paper. Code availability The code necessary for replication of the results in this paper is available for download at https://figshare.com/s/c984d29316e012b79e32. Articles Acknowledgements We acknowledge the World Climate Research Programme’s Working Group on Coupled Modelling, which is responsible for CMIP, and we thank the climate modelling groups for producing and making available their model output. We are grateful for U. Beyerle and J. Sedlacek at ETH Zurich, who provided access to the CMIP data. This research was supported by funding from the California Strategic Growth Council Climate Change Research Program (no. CCRP0056). P. Stainier provided valuable research assistance. Author contributions A.B. and J.S. designed the research, interpreted the data and wrote the paper. A.B. analysed the data. NATUrE ClimATE CHAngE Competing interests The authors declare no competing interests. Additional information Supplementary information is available for this paper at https://doi.org/10.1038/ s41558-019-0632-4. Correspondence and requests for materials should be addressed to A.B. Reprints and permissions information is available at www.nature.com/reprints. Nature Climate Change www.nature.com/natureclimatechange Last updated by author(s): Oct 2, 2019 Reporting Summary Nature Research wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency in reporting. For further information on Nature Research policies, see Authors & Referees and the Editorial Policy Checklist. Statistics nature research reporting summary Corresponding author(s): Alan Barreca For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section. n/a Confirmed The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly The statistical test(s) used AND whether they are one- or two-sided Only common tests should be described solely by name; describe more complex techniques in the Methods section. A description of all covariates tested A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient) AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals) For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted Give P values as exact values whenever suitable. For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated Our web collection on statistics for biologists contains articles on many of the points above. Software and code Policy information about availability of computer code Data collection Weather and birth records were downloaded from publicly available websites using "wget" command version 1.12 within Unix. The data were collapsed down to the county-day level using the standard "collapse" command in Stata-MP version 15.1. The climate change projections were compiled in R using "nc_open" and "ncvar_get" commands. Data analysis The main analysis was conducted using the standard "reghdfe" command in Stata-MP version 15.1. For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors/reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Research guidelines for submitting code & software for further information. Data Policy information about availability of data All manuscripts must include a data availability statement. This statement should provide the following information, where applicable: - Accession codes, unique identifiers, or web links for publicly available datasets - A list of figures that have associated raw data - A description of any restrictions on data availability October 2018 All data and code necessary for replication of the results in this paper are available for download at https://figshare.com/s/c984d29316e012b79e32 . The original weather data can be downloaded from ftp://ftp.ncdc.noaa.gov/pub/data/ghcn/daily . The original birth records can be downloaded from http://www.nber.org/ natality/ . The population data can be downloaded from https://seer.cancer.gov/popdata/download.html . Interested researchers should contact Urs Beyerle or Jan Sedlacek at ETH Zurich to gain access the CMIP5 climate projection data. 1 Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection. Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf Behavioural & social sciences study design All studies must disclose on these points even when the disclosure is negative. Study description Quantitative methods were used to analyze the correlation between atypical daily birthrates and temperature patterns in the United States between 1969 and 1988. Research sample The sample includes the near universe of births in all counties in the continental United States between 1969 and 1988, the period in which day of birth is available in the publicly available birth records. The United States Vital Statistics used a 50% sample in the earlier years of our sample period and slowly rolled out to a 100% sample across all states by the end of our sample period. Sampling strategy We used all available states in the continental United States. Data collection Not applicable. Data were collected by the National Center for Health Statistics. Timing Cohorts span the 1969-1988 period. Data exclusions Hawaii and Alaska were excluded due to having more unusual climates. Non-participation Not applicable. Randomization Not applicable. nature research reporting summary Field-specific reporting Reporting for specific materials, systems and methods We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. Materials & experimental systems Methods n/a Involved in the study n/a Involved in the study Antibodies ChIP-seq Eukaryotic cell lines Flow cytometry Palaeontology MRI-based neuroimaging Animals and other organisms Human research participants Clinical data October 2018 2