2.1. What is heat exposure?

Heat exposure it often discussed interchangeably with air temperature – that is, the temperature that can be measured with a standard thermometer. However, while air temperature is a major contributor to the level of heat that a person experiences, there are additional contributors. As McGregor and Vanos [21] explain in their primer on the physiological impacts of heat on the human body, these include other weather conditions (humidity, windspeed, radiation), individual factors (levels of exertion, pre-existing medical conditions, and medications), and the built environment.

Many researchers argue that humidity is particularly important [15,17,22]. This is because one way that the body cools itself is through having sweat evaporated from the skin, and sweat evaporates more slowly when it is humid. Therefore, for a given temperature, higher humidity increases the body’s risk of overheating.

In deciding how to measure heat exposure, we must understand which aspects – such as air temperatures, humidity or other contributors – matter for prenatal development. The difficulty here is that there are multiple mechanisms through which heat exposure during pregnancy can affect the mother and developing fetus [23], and hence lead to poorer birth outcomes. We do not yet know from physiological science which of these mechanisms is most important [24]. In the remainder of this section, we explain these mechanisms (summarised in Fig 2), before returning to the question of which measures may best capture these effects.

2.2. How does heat exposure affect the fetus?

We set out three channels through which heat exposure may affect prenatal development: heat stress, heat strain and maternal behaviour.

It is important to note that while these mechanisms can increase the risk of adverse birth outcomes like preterm birth – which is the focus of this study – they will not necessarily do so. They may affect fetal development in ways that impact a child’s health and development in infancy and/or later in life, but which do not necessarily lead to preterm birth.

2.3. Implications for heat exposure measurement

Given the mechanisms outlined above, we draw three takeaways for how we might construct heat exposure metrics.

First, it is not obvious which underlying measure of heat we should use. Within the literature, heat exposure is almost universally measured based on air temperature. But as described above, a person’s experience of heat is a result of a range of factors, including temperature, humidity, wind and radiation. Therefore, in theory, a data series incorporating these variables may provide a more accurate picture of true heat exposure. One option is wet bulb temperature, which can be observed directly, and reflects air temperature and humidity – other options include Apparent Temperature, Heat Index, Universal Thermal Climate Index, and Wet Bulb Global Temperature [15]. However the reason that this is not an obvious choice is that, empirically, we do not see a consistent link between high humidity and poorer health outcomes. As Baldwin et al. [16] suggest, there may be a range of explanations for this, including the fact that a build-up of humidity when it is not raining has different effects than high humidity while raining.

Second, that the effects of extreme heat – that is, heat at levels around or above normal core body temperatures – may be different from the effects of moderate heat, because extreme heat is more likely to cause heat stress. This means estimates of the impact of heat exposure based on average temperatures over a given period are likely mis-specified, as an average will fail to distinguish between conditions likely to lead to heat strain vs heat stress.

Third, that even moderate levels of heat, when experienced for a prolonged period, may affect fetal development by causing heat strain or changes in maternal behaviour. If this is the case, then it may be important to choose metrics that reflect prolonged heat exposure – such as high daily minimums, reflecting at least 24 hours of high temperatures – in addition to the metrics used in the overwhelming majority of studies which measure peak levels of heat exposure (i.e., high daily maximums). This may be particularly important in tropical climates where daily maximums and minimums are not strongly correlated with one another (as shown in Fig 1).

In our analysis, we estimate the impact of heat exposure on preterm birth, using a range of different heat metrics, to test which fits the data best. In line with this conceptual framework, we test metrics that include both air temperature and wet bulb temperature, metrics based on daily maximums and minimums, and metrics allowing for non-linear effects of heat exposure. We compare these metrics with those most commonly used in prior research (described below). However, our conceptual framework also points out a potential moderator between heat exposure and birth outcomes that we are unable to measure in our data – maternal behaviour. While maternal behavioural responses to heat exposure do not feature in our data analysis, behaviour is likely a key factor, and may help to explain differences in the magnitudes and timing of effects (in terms of whether exposure has different effects depending on the timing of exposure during pregnancy) across different populations and contexts.

3.1. Study context

The Northern Territory (NT) is one of Australia’s eight states and territories. The NT is a large region covering the central part of northern Australia, in which residents face high levels of heat exposure. The NT has a population of around 233,000, 60 percent of whom live in or around the capital city of Darwin [36]. Previous research shows that heat exposure plays a major role in explaining month-to-month variation in average birth outcomes in the NT [37].

The tropical north of the NT, where Darwin is located, is hot and humid. Temperatures vary within season but, on an average day in the wet season (November to April), temperatures range between 25–33 degrees Celsius, and in the dry season (May to October), between 20–30 degrees. The hottest time of year is October-December, in the ‘build up’ to the wet season – when temperatures and humidity are high, and there is little rainfall. Heavy rainfall usually begins in late December, though the timing can vary from year to year.

The central and southern parts of the NT have an arid climate, with very hot summers and mild or cold winters. In Alice Springs, the largest town in the region, temperatures on an average summer day range between 20 and 35 degrees, and on an average winter day, between 4 and 20 degrees.

Around one-third of the NT population identify as Aboriginal. There are significant differences between the Aboriginal and non-Aboriginal populations in the NT, in terms of geography, heat exposure and economic resources. Eighty percent of Aboriginal residents in the NT live outside of Darwin, many in remote Aboriginal communities which experience more extreme weather conditions. In remote communities, many houses are poorly insulated, and many residents face energy poverty [38].

3.2. Data

We combine administrative birth records with daily weather data.

Our analysis sample includes 95 percent of all babies who were born in the NT and conceived between March 2000 and September 2009, a sample of 34,258 children. There were 35,899 babies born during this period. Our sample excludes 1,020 births for which the mother’s place of residence could not be geocoded and therefore could not be matched to weather data, and 621 births for which some covariates were missing. We define our sample based on date of conception instead of date of birth because definitions based on date of birth will systematically exclude children born preterm at the beginning of the sample period, and exclude those born late term at the end of the sample period – this could lead to bias, especially when analysing the impact of seasonal exposures like heat [39]. We determine date of conception by subtracting gestational age (in weeks) from birthdate.

We include only births to mothers whose usual place of residence is in the NT, and could be geo-coded. There are a small number of births for which the mother’s place of residence as entered in the perinatal data could not be found, either using a fuzzy match with the R package ‘geonames’, or through manual search on Google, the NT Place Names Register (https://www.ntlis.nt.gov.au/placenames/) and BushTel (https://bushtel.nt.gov.au/). We include stillbirths (making up under 1% of births) and plural births (under 2.5% of births) in the analysis sample.

The birth record data were extracted from administrative records and prepared for analysis between 3 July and 31 August 2017. Data analysis for research purposes for this paper was conducted between 26 June 2023 and 19 December 2024 using Stata 17. Analysis data are de-identified, and the authors do not have access to information that could identify individual participants during or after the data were collected. Approval for the research has been provided by the Northern Territory Department of Health and Menzies School of Health Research Human Research Ethics Committee (Ref: 18–3261). The data linkage project’s First Nations Advisory Group have independently reviewed and endorsed the research to ensure it is respectful of Aboriginal perspectives.

We link these records to NASA’s daily weather reanalysis data, based on the mother’s place of usual residence at the time of the birth. We do not have data on place of residence throughout pregnancy, therefore we assume that the place of residence at birth corresponds with the place the mother spent most of her time during pregnancy. The weather data are available at intervals of 0.5 x 0.625 degrees of latitude and longitude (roughly 50x55km). This gives a total of just over 500 cells throughout the NT, 166 of which have births within our analysis period – an average of 203 births per cell. We link data based on date of conception, such that the first trimester is the first 12 weeks from the conception date, the second trimester is the following 14 weeks, and the third trimester is the following 13 weeks. Definitions for timing and length of trimesters vary across contexts, but these are the definitions commonly used in Australia. We analyse 39 weeks of gestation, as this is the average gestational length in our population.

The weather reanalysis data comes from NASA’s model using ground station and satellite data. Some may be concerned about accuracy of these data compared with traditional weather station observations. However, in our context, the NASA data lines up very closely with observations from the Australian Bureau of Meteorology’s weather stations (see S1 Appendix for details). We use the reanalysis data because they are highly localised and have no missing observations – which is not the case for weather stations across remote parts of the NT. For instance, over this period some weather stations are decommissioned, and others have systematic missing observations (i.e., in the wet season when they become inaccessible due to flooding and therefore observations are systematically missing during periods of extreme weather).

3.3. Heat exposure metrics

We construct five different heat exposure metrics. The first three are based on common metrics used in prior research, while the next two build on the conceptual framework outlined above. Our metrics are:

Wet bulb.

This metric is analogous to the ‘Max and min’ metric described above, but using wet bulb temperatures as the underlying data series, instead of air temperature. We construct counts of the number of days within each trimester with daily average wet bulb temperatures with ranges of <10, 10–14.99, 15–19.99, 20–24.99, 25 + , and the number of days with daily maximum wet bulb temperatures within ranges of <10, 10–14.99, 15–19.99, 20–24.99, 25–29.99 and 30 + . This metric reflets the possibility highlighted in our Conceptual framework that wet bulb temperatures may measure heat exposure more directly than air temperatures.

Maximum wet bulb temperature is not available directly within the NASA data, but we construct this measure using daily maximum combined with relative humidity, using Stull’s equation [56]. We use average wet bulb temperature instead of minimum because this measure is directly available. However, as daily average is a linear transformation between maximum and minimum, the relationship between maximum and minimum (for air temperature), and maximum and average (for wet bulb temperature) should be constant.

Table 1 sets out the mean, standard deviation, maximum and minimum of these measures over the full pregnancy, for babies in our sample. During an average pregnancy, there are 2.4 heatwaves and just over one week with maximum temperatures above 40 degrees – but this varies greatly, with some pregnancies experiencing up to 9 heatwaves, and 98 days with maximum temperatures above 40 degrees.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics.

https://doi.org/10.1371/journal.pone.0330498.t001

3.4. Outcome measure

Our main outcome measure is preterm birth, defined as birth before 37 complete weeks of pregnancy. In our population, 10 percent of babies are born preterm (Table 1). We focus on preterm birth for comparability with previous research: it is the most commonly studied birth outcome. However, our conclusions hold across four additional measures of health at birth: birthweight, small for gestational age, Apgar scores, and admission to a special care nursery. Estimates using these outcomes are presented in the supplementary material.

Preterm birth is an important intermediary outcome: a large body of research tells us that children born preterm are more likely to face worse health and educational outcomes [19,57,58]. In the NT, children who were born preterm are more likely to be assessed as developmentally vulnerable at age 5 [59].

Recent research demonstrates that this relationship is not fixed: the predictive power of preterm birth has declined over time [60], and many children born preterm may face no detectable long-term effects [61]. This makes sense, for two reasons. First, advances in neonatal healthcare have greatly improved outcomes for preterm-born infants [62]. Second, because preterm birth has many causes [63]. The aetiology of preterm birth, and hence the long-term outcomes associated with it, may differ over time and across contexts. However, even in the desirable situation where preterm birth does not increase the risk of long-term developmental vulnerabilities for an individual, it remains an outcome of interest given the high cost of providing remedial neonatal care [20].

3.5. Analytical methods

Our goal in this analysis is to assess whether and how our estimates of the causal impact of heat exposure on preterm birth change when we use different heat exposure metrics. A challenge in doing so is that heat exposure may be correlated with birth outcomes for reasons which do not necessarily reflect the causal impact of heat itself.

Of primary concern is omitted variable bias. Heat exposure varies both over time, and across locations. But most of the variation in birth outcomes over time and across locations is not due to heat exposure. For instance, socioeconomic status is known to affect birth outcomes, but is also likely to affect both place of residence and timing of conception [64]. A link with place of residence may be particularly important in the NT, given that most non-Aboriginal people live in the capital city of Darwin, whereas most Aboriginal people live in remote communities [65], many of which face higher levels of heat exposure than Darwin. Similarly, we know that there are seasonal risks in addition to heat exposure, such as disease prevalence and economic conditions, which contribute to month-to-month variations in birth outcomes [37].

Our analytical approach is to estimate the causal effect of heat exposure by isolating fluctuations in heat that are exogenous to these (possibly endogenous) sources of time- and place-based variation. Exposure to such exogenous variation is beyond individuals’ control, and cannot be anticipated more than a few days in advance. This variation in exposure is, therefore, as good as random.

To do this, we specify a regression model with fixed effects for time and place. This approach is discussed in detail in Dell et al.’s review [40], and has been widely used in prior research.

As is common practice, we allow for the effects of heat exposure to vary by trimester; this allows for the possibility, as discussed in the Conceptual framework, that the effects of heat exposure may depend on its timing – for instance, exposure in the first trimester may increase the risk of pre-eclampsia [27], and exposure in the third trimester may bring forward labour [33].

We therefore estimate the following linear fixed effects model:

(1)

where is an indicator of whether baby i, who was conceived on date t (in month m, year y), was born preterm or not. k is an index of the three trimesters of pregnancy. X is a set of individual-level covariates, which includes an indicator for whether it is the mother’s first pregnancy, the age of the mother (in 5-year age bands), and an indicator for whether the baby is Aboriginal. Standard errors are clustered at the location level. We identify location, j, based on the mother’s suburb/community of residence, which we have geocoded, and grouped into locations of maximum 50-mile distance between points using cluster analysis. There are 106 such clusters in our data, with an average of 311 births per cluster.

Heat represents any one of the five heat metrics outlined above. They are based on daily weather observations from small-area location j, which are summarised over each trimester based on exact date of conception t. Our coefficients of interest are .

In specifying the fixed effects, we follow prior studies in using fixed effects for the month-year of conception, , and for the mother’s place of residence, interacted with the month of conception and the baby’s sex, . Interacting location fixed effects with month and sex allows for the possibility that there may be different seasonal patterns in birth outcome across locations – this may be the case in the NT given, for example, that some communities experience regular flooding in the wet season, and hence their experience of the wet season may be different from other communities that do not flood. Furthermore, we know that different types of exposures may affect male and female fetuses differently [66]. Interacting these location- and season-specific effects with sex allows for this possibility.

To illustrate what we are comparing with these community-month-sex fixed effects, consider two boys born in Darwin in January of the same year. Both boys were conceived in the same location (Darwin) and the same month (April), so they share the same location-month-sex fixed effect, though they were conceived two weeks apart. They also share the same month-year fixed effects. However, their mothers experienced different heat exposure during pregnancy due to particularly warm weather occurring in January while one baby remains in utero but the other has already been born. The fixed effects allow us to compare outcome for these two boys, attributing any difference to the variation in heat exposure their mothers experienced, while holding constant all other factors that vary systematically by location, month of conception, and baby’s sex.

This day-to-day variation in weather within the same location-month-sex group is the source of identification in our model. It represents truly exogenous variation—mothers cannot anticipate or control whether their pregnancy will coincide with an unseasonably hot days versus a cooler days within the same general season and location.

Our choice of interacted fixed effects follows prior studies [4,67]. That said, S1 Table shows that our estimates are similar regardless of the specifics of how we define these fixed effects.

3.6. Our approach to selecting a preferred heat metric

Part of our goal in this analysis is descriptive: to run the same model sequentially with each heat metric, and learn how choice of heat metric affects our conclusions around the impacts of heat exposure on preterm birth. To do this, we present regression coefficients for the same model, using each of our five heat metrics.

However, we also wish to select a preferred heat metric. Out of our five candidate metrics, we want to know which one best captures the impact of heat exposure on preterm birth in the NT. In doing so, we face a challenge: because each regression uses different heat exposure metrics, the magnitudes and statistical significance of the resulting coefficients cannot be directly compared with each other.

We therefore take a two-stage approach to selecting a preferred metric. First, we present an F-test of joint significance of the regressors included in each heat exposure metric. If the regressors are not jointly statistically significant, we exclude the metric from further comparison. Second, we compare goodness of fit, for the remaining metrics.

We use two measures of goodness of fit. First, the R-squared. This tells us how much variation in the outcome is explained by the regressors. A standard R-squared increases when additional regressors are added. To account for different numbers of regressors in each metric, we use an adjusted R-squared: this penalises additional regressors. A higher adjusted R-squared, therefore, indicates that the model fits the data better, when comparing among models with different numbers of regressors. Second, we present the Akaike Information Criterion (AIC). The AIC helps us to compare several alternative, non-nested models, on a single measure: it is calculated as the likelihood of the model (estimated using maximum likelihood), penalised for the number of regressors. It therefore weighs up model fit with model complexity, allowing us to compare multiple working hypotheses against each other [68]. In their review on variable selection, Heinze, Wallisch and Dunkler [69] recommend using AIC in cases such as ours, where theory supports a relatively small set of competing models, and we want to select between them. In comparing AIC scores, a lower score indicates better fit.

4.1. Goodness of fit comparison

Table 2 presents a comparison of how well each heat metric fits the data.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Measures of model fit for each heat metric.

https://doi.org/10.1371/journal.pone.0330498.t002

Across our five metrics, we can first exclude ‘Trimester average’ and ‘Heatwave count’ from comparison: in an F-test of joint significance at the 5% level of significance, we would fail to reject the hypothesis that the coefficients in these metrics are jointly equal to zero, hence that they do not help to explain variation in preterm birth rates.

We then have three remaining metrics. Across all metrics, the adjusted R-squared is low and varies little, ranging between 0.035 to 0.037. However, the metrics produce very different AIC scores. On both R-squared and AIC measures, the ‘Max and min’ metric performs better: it has a slightly higher adjusted R-squared and substantially lower AIC.

We separately consider whether a combination of metrics provides a better fit – for instance, combining the ‘Max and min’ metric, the ‘Wet bulb’ metric and ‘Heatwave count’ metric together. However, while some of the additional coefficients are statistically significant, they do not meaningfully improve model fit (S2 Table), and interpretation of individual coefficients becomes very difficult.

We therefore select the ‘Max and min’ metric as our preferred heat exposure metric.

4.2. Coefficient estimates

Having identified the ‘Max and min’ metric as our preferred one, we now turn to a presentation of the coefficients for the regressors in each heat metric. This tells us both the estimates from our preferred model, and the potential consequences, in terms of inferences we may draw, from selecting a non-preferred metric.

We present estimates graphically, showing coefficients and their 95% confidence intervals. Tables with all coefficients cited and their standard errors can be found in S3–S10 Tables.

Benchmark heat metric.

Fig 3 presents estimates from our benchmark metric. We can see little impact from heat exposure in the first trimester. However, in the second and third trimesters we find that exposure to cooler temperatures (below our omitted category of 25–30 degrees) contributes to lower risk of preterm birth; we see no additional impact of heat exposure above 30 degrees. This suggests that it may be predominantly heat strain (i.e., prolonged exposure to moderate heat), instead of heat stress (i.e., exposure to extreme heat) affecting pregnancies in the second and third trimesters.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Benchmark heat metric.

This figure shows the regression coefficients and 95% confidence intervals from Equation 1, where the heat metric is a count of the number of days within each trimester that the daily maximum temperature fell into the following ranges: < 20, 20–25, 25–30, 30–35, 35–40, 40 + . The omitted category is 25–30 degrees. Sample size: 34,258.

https://doi.org/10.1371/journal.pone.0330498.g003

If it is the case that maternal heat strain affects the baby through reduced flow of nutrients to the placenta [24], this could explain why we see positive effects of cooler temperatures in the second and third trimesters, but not the first trimester: in the first trimester, the fetus receives nourishment from the yolk sack and not the placenta, and is therefore less reliant on maternal blood flow for nutrients. When the placenta takes over around the beginning of the 2^nd trimester, this is when we might expect an impact of heat strain.

Average temperatures.

Fig 4 shows estimates from the ‘average temperature’ metric. Note the difference in scale from Fig 3 due to differences in units: a 1 degree increase in average temperatures represents more heat exposure than a single day of extreme heat. For example, in a trimester with average temperatures of 30 degrees, a 1 degree increase in the average would represent over one week of temperatures of 40 or above.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Trimester average.

This figure shows the regression coefficients and 95% confidence intervals from Equation 1, where the heat metric the trimester-average of daily maximum temperatures. Sample size: 34,258.

https://doi.org/10.1371/journal.pone.0330498.g004

Fig 4 shows that a 1 degree increase in average temperatures in the third trimester leads to a 0.007 percentage point increase in the probability of preterm birth. Although we saw some statistically significant impacts of heat exposure in the first and second trimesters of the benchmark model, these are not evident here.

Heatwave count.

Fig 5 shows estimates from our ‘heatwave count’ metric. We find that while the coefficients are positive (indicating that experiencing an additional heatwave in utero increases the risk of preterm birth), they are not statistically significant. Though this is not the case for all outcomes – we find statistically significant coefficients on average birthweight, with each additional heatwave in the second trimester associated to a 5.7 gram reduction (S5 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Heatwaves.

This figure shows the regression coefficients and 95% confidence intervals from Equation 1, where the heat metric is a count of the number of heatwaves. Heatwaves are defined as 3 or more consecutive days of unusually hot weather relative to usual conditions within a location – using on the Australian Bureau of Meteorology’s method. Sample size: 34,258.

https://doi.org/10.1371/journal.pone.0330498.g005

The fact that we do not find statistically significant effects on preterm birth makes sense given that, in our benchmark model, much of the effect of heat exposure came through improved outcomes from exposure to cooler temperatures (below 25–30 degrees), and not from extreme heat conditions that would contribute to a heatwave.

Max and min (preferred metric).

The ‘Max and min’ metric – our preferred metric, based on goodness of fit – adds minimum temperature exposure bands to our benchmark metric. We see similar patterns to the benchmark metric, though the addition of minimum temperatures reveals further effects of prolonged heat exposure – i.e., increased risk of preterm birth for babies exposed to more days when minimum temperatures do not fall below 20 degrees (Fig 6).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Max + min.

This figure shows the regression coefficients and 95% confidence intervals from Equation 1, where the heat metric is a count of the number of days within each trimester that the daily maximum temperature fell into the following ranges: < 20, 20–25, 25–30, 30–35, 35–40, 40 + , and a count of the number of days within each trimester that daily minimum temperatures fell into the following ranges: < 5, 5–10, 10–15, 15–20, 20–25, 25 + . The omitted categories are 25–30 degrees (max) and 15–20 (min). Sample size: 34,258.

https://doi.org/10.1371/journal.pone.0330498.g006

In the first trimester, high minimum temperatures of 20 or above increase the risk of preterm birth, while low minimums (below 10 degrees) reduce the risk; we no longer see a statistically significant impact of high maximum temperatures.

In the second trimester, we continue to see a protective impact of low maximum temperatures (below 25 degrees), and these estimates are larger in magnitude here than in the benchmark metric. We also find that high minimum temperatures (25 degrees or above) increase the risk of preterm birth.

In the third trimester, we now see that high maximum temperatures (between 30–39.99 degrees) increase the risk of preterm birth, though with no statistically significant impact of minimum temperatures. Some of these changes in estimates may be due to multicollinearity, given that low values of maximum and minimum temperatures each have similar coefficients in the third trimester when modelled separately (S3 Table).

As this is our preferred metric, we ran our analysis separately for Aboriginal and non-Aboriginal babies, to assess whether the effects of heat exposure are different amongst these different populations (see S7 Table). We find estimates are similar in both groups, though among Aboriginal children, the only statistically significant estimates are for the third trimester: exposure to cooler maximum temperatures reduces the risk of preterm birth, and exposure to temperatures of 35–40 degrees increases the risk. In contrast, the effects of higher maximum and minimum temperatures in the first and second trimesters are statistically significant among non-Aboriginal children, but there is little effect of third trimester exposure. It may be that Aboriginal communities and cultural practices are better adapted to hot weather, leading to little impact of a marginal increase in heat exposure. It may also be the lack of statistically significant effects reflects the greater diversity among Aboriginal mothers in their exposures to heat due to the varied locations of Aboriginal communities – leading to estimates that are generally of similar magnitude to those for non-Aboriginal mothers, but less precisely estimated [37].

Wet bulb.

Fig 7 shows coefficients for analysis using our ‘Wet bulb’ metric – which is analogous to the ‘Max and min’ metric, but constructed using wet bulb temperatures instead of air temperature. We see little statistically significant relationship between wet bulb temperatures and preterm birth: the only statistically significant coefficient suggests that high wet bulb temperatures (25 to 29.99 degrees) in the third trimester decrease the risk of preterm birth – that is, an effect in the opposite direction to that suggested by all other metrics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Wet bulb.

This figure shows the regression coefficients and 95% confidence intervals from Equation 1, where the heat metric is a count of the number of days within each trimester that the daily maximum wet bulb temperature fell into the following ranges: < 10, 10–15, 15–20, 20–25, 25–30, 30 + , and a count of the number of days within each trimester that daily average wet bulb temperature fell into the following ranges: < 10, 10–15, 15–20, 20–25, 25 + . The omitted categories are 20–25 (max) and 15–20 (avg). Sample size: 34,258.

https://doi.org/10.1371/journal.pone.0330498.g007

Fig 7 presents both average and maximum daily wet bulb temperatures in a single metric. But sensitivity testing shows that our conclusions would be similar if we were to construct our metric using maximum or average wet bulb values separately. As shown in S9–S10 Tables, we find small statistically significant coefficients on low values of daily maximum wet bulb temperatures in trimesters 1 and 3, but this model has low predictive power.

4.3 Do these findings differ by climate zone?

As described in the Introduction, the relationship between various measures of heat exposure – and hence the extent to which selection of heat metric matters – varies across climate zones. We therefore repeat our analysis, but running models separately for the tropical climate zone in the north of the NT, and the arid climate zone in central NT.

We see some differences in coefficients between the climate zones (S11 Table). For instance, while high minimum temperatures in both the first and second trimesters significantly increase the risk of preterm birth in the tropical climate zone, we do not find this in the arid climate zones. In addition, while high maximum temperatures increase the risk of preterm birth in both climate zones, the magnitudes of these coefficients are at least twice as large in arid climate zones.

We may have expected the metric based on wet bulb measures to be a better fit in the tropical climate zone. However, we do not find this (S12 Table). Across both climate zones, the ‘Max and min’ metric provides the best fit (S13 Table).

5.1. Practical implications of using non-preferred heat metrics

The analysis we have presented above demonstrates the statistical significance and model fit of five alternative metrics. But the varied scales of the coefficients make it difficult to compare the practical significance of these different estimates. While we do find one metric that outperforms the alternatives on model fit, the differences appear marginal – our preferred ‘max and min’ metric has an adjusted R-squared that is just 0.001 units higher than the benchmark metric. Furthermore, the explanatory power of even our preferred heat metric is low – clearly there are many more proximate factors, not included in our models, that explain the vast majority of variation in preterm birth rates. That said, while the effects of prenatal heat exposure are small for any one individual, these impacts are important at the population level. This is particularly so in the context of high and rising levels of heat exposure.

To quantify the implications of metric choice, we use our regression estimates to generate predictions of the impact of typical within-year differences in heat exposure on preterm birth rates.

We generate two sets of predictions: first, the hypothetical case where all babies in our analysis data were exposed to heat typical of the hottest 9 months of the year in utero (born in May), and second, the hypothetical case where all babies were exposed to heat typical of the coolest 9 months of the year (born in November). We then compare the difference between these two predictions in Fig 8. Our methodology is outlined in S2 Appendix.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Model predictions with 95%, 80% and 68% confidence intervals.

Predictions shown for all heat metrics, and represent the models’ predicted rate of preterm birth for a population exposed to typical conditions in the coolest 9 months of the year in utero, compared with a population exposed typical conditions in the hottest 9 months of the year in utero. The confidence intervals are shown around the estimate for the coolest 9 months for each metric.

https://doi.org/10.1371/journal.pone.0330498.g008

In our preferred ‘Max and min’ heat metric, we estimate that if all babies were in utero during the coolest 9 months of the year, preterm birth rates would be 9 percent, and if all babies were in utero during the hottest 9 months of the year, preterm birth rates would be 13.5 percent. This is a 4.5 percentage point difference.

In contrast, the alternative metrics imply effects of less than half this size: we estimate a 2 percentage point difference using the benchmark metric, and smaller differences from the other metrics. Therefore, choosing a non-preferred metric means not only that we may mis-understand the relationship between heat exposure and preterm birth, but also that we may severely under-estimate heat’s implications for population health.

Of course, there is uncertainty around these predictions: the difference between these two predictions with our preferred metric is not statistically significant at the 95% level. However, based on the large body of research and our own analysis, we know that it is more likely than not that hot weather causes a substantial worsening in health at birth, and a 4.5 percentage point effect is our best estimate of the impact of normal seasonal variation in the NT. To put this in context, an effect of this magnitude is equivalent to intensive rates of smoking during pregnancy [18].

5.2. Recommendations for future analysis

In our analysis, we have compared five different approaches to defining and measuring heat exposure – we call these ‘heat metrics’. We have found that a metric based on air temperature which allows for non-linear effects of increased temperatures, and includes both daily minimum and maximum temperatures, provides the best fit. Our findings lead us to two general recommendations for future research.

First, it is very unlikely that the true relationship between heat exposure and birth outcomes is the one implied by either the ‘Trimester average’ or the ‘Heatwave count’ metric. As outlined in our Conceptual framework, we have good reason to believe that both moderate and extreme heat during pregnancy can affect birth outcomes, possibly in different ways. Therefore, at a minimum, our models should allow for these nonlinear effects. While some studies have found that the ‘Trimester average’ and ‘Heatwave count’ metrics nevertheless produce statistically significant estimates, we have found that they will not necessarily do so. Furthermore, as shown in Fig 8, even among alternative metrics where the regressors are statistically significant, the implied impact of heat exposure from a more limited metric may differ from a more flexible metric by orders of magnitude. Therefore, where researchers have the choice, we recommend against using the trimester average or heatwave count metrics.

Second, in the population we analyse, we find that metrics based on air temperature, and not on wet bulb temperatures, better capture the effects of heat exposure – even though most of this population lives in a humid, tropical climate zone. While surprising, our findings are not unique; many epidemiological studies have also found little impact of humidity on health outcomes [16]. In understanding this finding, we cannot rule out issues with data quality and choice of measure. Our wet bulb data are derived from satellite images, and not directly measured at a weather station. This is the only metric incorporating humidity that is available within our region – weather station data do not cover the whole of the NT, meaning we do not have the underlying data series needed to construct alternative heat indices. It may be that, if they were available, alternative heat indices that combine heat and humidity could have more explanatory power than wet bulb measures. That said, we note that studies using reanalysis wet bulb data for other outcomes, such as LoPalo’s [71] analysis of the impacts of heat on worker productivity, do find wet bulb measures to be more predictive than ambient temperatures.

However, our findings may also reflect real, complex relationships between heat, humidity and health in a tropical climate. The role of climate may explain, for instance, why a recent study from Western Australia finds clear effects from a metric based on a heat index combining temperature and humidity [72]. While some parts of Western Australia have a tropical climate, the vast majority of the population in that study lives in Perth, which has a dry, Mediterranean climate, and hence humidity will play little role in determining variation in the index.

It may be that the interplay of heat and humidity is not well captured by a single data series, whether it is wet bulb or another heat index. In Darwin, for instance, days with higher humidity and rainfall typically do not reach particularly high maximum temperatures; in fact, humidity and air temperatures are negatively correlated with each other (see S14 Table and S3 Fig). But because humidity is a contributor to wet bulb measures, wet bulb temperatures are mechanically higher on cooler, rainy days. Therefore, it is possible that hot, humid days with higher wet bulb temperatures can, in theory, contribute to greater heat strain. But the same wet bulb temperatures can also be the result of very humid, cooler-than-average and rainy days, which provide some relief from high temperatures. This may explain why our estimates suggest that higher wet bulb temperatures in the third trimester can actually reduce preterm birth rates. It may also be that the same level of humidity feels different on a day with intense rainfall (i.e., the wet season) compared with days with no rainfall (i.e., the build up to the wet season). Future research could analyse air temperature, humidity and rainfall, and the interaction between the three as separate regressors, to test whether this yields more informative estimates.

Our takeaway is that in the NT, as well as in many other contexts where the only data available are satellite or reanalysis data, heat exposure metrics based on air temperature may well provide more explanatory power than metrics based on wet bulb temperatures. However, future research could compare a range of heat indices in a location where data are known to be of high quality and where humidity and extreme heat co-occur: this would help to more definitively assess whether the lack of consistent effects of humidity that we and others find is due to data quality, or to a mis-specification of the relationship between heat, humidity and health.