Study population
Data were collected from all singleton fetal death and birth records in California (1996–2017), Florida (1991–2017), Georgia (1994–2017), Kansas (1991–2017), New Jersey (1991–2015), and Oregon (1991–2017). Years of data differed based on availability of fetal death records. Stillbirth was defined as a fetal death that occurred after 20 weeks gestation and was issued a fetal death certificate. Information collected from the fetal death and birth records included date of delivery, estimated gestational age, maternal age, education, race, ethnicity, and county of residence, Stillbirths and live births were excluded from the current study if gestational age and last menstrual period (LMP) date were missing or if the recorded gestational age was not between 20–44 weeks. To avoid the fixed cohort bias, which can occur when the study population is defined by birth date, we defined the study population based on LMP date and limited our sample to women whose LMP was between September 1st of the year prior to data availability (20 weeks prior to first possible birth) and February 28thof the last year of data availability (44 weeks prior to last possible birth), and did not restrict analyses to the warm season [14]. For example, in California, where data were available from 1996–2017, we included all births with an LMP date between September 1, 1995 – February 28, 2017.
Study design
We implemented a matched case–control design because of concerns about confounding by seasonal patterns of conception [15]; our internal simulations indicated that a case–control approach with matched gestational timing of the exposure windows would be less susceptible to this bias than a case-crossover approach. Stillbirths were matched 1:4 to live births on maternal race/ethnicity (white, non-Hispanic; black, non-Hispanic; Hispanic; other; missing), maternal education (less than high school; high school degree; some college; college degree or more; missing), and county. If four controls were not available, cases were matched to as many controls as possible. The exposure window for the stillbirths was the 6 days prior to delivery and the date of delivery (lag 0–6). The exposure window for the controls was the gestational week corresponding to the matched stillbirth’s exposure window, calculated by adding the case’s gestational age to the LMP of the matched control. Meteorology was assigned by county, a matching factor, such that estimated associations would be driven by temporal contrasts of exposure. In Oregon, maternal race information on the fetal death records was not included in the data transfer; therefore, cases and controls were not matched on race/ethnicity, and race/ethnicity was not adjusted for in the Oregon analysis.
Meteorologic data
Meteorologic data were collected for 1991–2017 from Daymet [16]. Daymet is a well-tested gridded meteorology dataset that uses ground-based in situ station observations and a collection of interpolation and regression algorithms to produce 1 km x 1 km gridded estimates of daily temperature and moisture, among other variables [17,18,19]. County-level temperatures were calculated by using the unweighted average of all grid cell estimates within a county. County-level temperature information was then linked to fetal death and birth records based on the reported maternal county and exposure window.
Exposure definitions
Heatwave definitions
Heatwaves were defined based on the mean daily temperature using the relative temperature threshold framework, where a hot day was defined as any day that the mean temperature was above a given threshold, which for this study was the county-specific 97.5thpercentile. The thresholds were county-specific because of climate differences both within and across states, and were defined using our full data period (1991–2017). Three heatwave definitions were created using the number of hot days in the previous week and the temperature over the threshold during the exposure window (lag days 0–6) [20]. Heatwave definition 1 (HW1) was a measure of the total number of hot days in the previous week, categorized as 0, 1, 2, and ≥ 3. Heatwave definition 2 (HW2) aimed to measure the impact of sustained heat and was defined as the number of consecutive hot days in the previous week. Separate indicator variables were created for ≥ 2 consecutive, ≥ 3 consecutive, and ≥ 4 consecutive days (i.e., if an exposure window had ≥ 4 consecutive hot days, it would also have ≥ 2 and ≥ 3 consecutive hot days). HW2 was defined using indicator variables, as opposed to exclusive categories, to allow for comparisons to previous literature. The final operationalization, heatwave definition 3 (HW3), was a continuous measure that incorporates both duration and intensity of hot days in the previous week, similar to an area under the curve measure. HW3 is the average difference between daily temperatures and the threshold during the exposure window; if the average was below the threshold, HW3 was given a value of 0.
Temperature definitions
In addition to the three heatwave definitions based on the temperature above a county-specific threshold, the 7-day average temperature was used to create one absolute measure and one relative measure of continuous temperature. The first was a measure of the 7-day average temperature during the exposure window; a categorical variable was also created with cut points to ensure sufficient sample size in each category (< 5 °C, 5–10 °C, 10–15 °C, 15–25 °C (REF), 25–27 °C, ≥ 27 °C). Second, because of the potential for acclimatization to local temperature norms, we assigned county-level percentiles to the 7-day average temperature based on the temperature data from 1991–2017; these were also categorized (< 2.5%, 2.5–10%, 10–25%, 25–75%, 75–90%, 90–97.5%, ≥ 97.5%).
Statistical analysis
State specific odds ratios (ORs), 95% confidence intervals (CI), and variance–covariance matrices were estimated using conditional logistic regression models adjusting for maternal age (10–19; 20–24; 25–29; 30–34; 34–39; ≥ 40 years), LMP month (to control for recurrent seasonal trends), and LMP year (to control for long term trends). HW1 was modeled as a categorical exposure (0 days (REF), 1, 2, ≥ 3 days). HW2 was estimated using three separate models with binary exposures (yes/no: ≥ 2 consecutive days; ≥ 3 consecutive days; ≥ 4 consecutive days). HW3 was modeled as a continuous exposure, estimating the OR corresponding to a 1 °C increase over the threshold in the previous week. Results were stratified by the timing of the stillbirth (early (< 28 weeks) v. late (≥ 28 weeks, including term stillbirths)) and maternal race/ethnicity (white, NH; black, NH; Hispanic; other) was assessed using stratified analyses. We chose to examine stillbirths based on gestational age at delivery because it is thought that early stillbirths are difficult to prevent without early intervention and are more commonly due to genetic abnormalities [2, 21]. Continuous temperature and percentile were modeled using natural cubic splines to allow for nonlinear patterns. Placement and number of knots were selected based on cut-off values for the categorical variables and the distribution of absolute temperatures across states. Knots for continuous temperature were placed at 5, 20, and 25 degrees; continuous percentile knots were placed at 2.5, 10, 25, 75, 90, and 97.5 percent. As a sensitivity analysis, categorical parameterizations of continuous temperature and percentile were used and results were compared to those from the spline models. State-specific estimates were combined using a multivariate fixed-effect meta-analysis to estimate the average association across states. Specifically, we calculated a weighted average of state-specific vectors of log odds where the weight corresponds to the inverse of their corresponding variance–covariance matrices. Statistical analyses were completed using SAS 9.4 and R.