High ambient temperature in pregnancy and risk of childhood acute lymphoblastic leukaemia: an observational study

Summary Background High ambient temperature is increasingly common due to climate change and is associated with risk of adverse pregnancy outcomes. Acute lymphoblastic leukaemia is the most common malignancy in children, the incidence is increasing, and in the USA disproportionately affects Latino children. We aimed to investigate the potential association between high ambient temperature in pregnancy and risk of childhood acute lymphoblastic leukaemia. Methods We used data from California birth records (children born from Jan 1, 1982, to Dec 31, 2015) and California Cancer Registry (those diagnosed with childhood cancer in California from Jan 1, 1988, to Dec 31, 2015) to identify acute lymphoblastic leukaemia cases diagnosed in infants and children aged 14 years and younger and controls matched by sex, race, ethnicity, and date of last menstrual period. Ambient temperatures were estimated on a 1-km grid. The association between ambient temperature and acute lymphoblastic leukaemia was evaluated per gestational week, restricted to May–September, adjusting for confounders. Bayesian meta-regression was applied to identify critical exposure windows. For sensitivity analyses, we evaluated a 90-day pre-pregnancy period (assuming no direct effect before pregnancy), adjusted for relative humidity and particulate matter less than 2·5 microns in aerodynamic diameter, and constructed an alternatively matched dataset for exposure contrast by seasonality. Findings 6849 cases of childhood acute lymphoblastic leukaemia were identified and, of these, 6258 had sufficient data for study inclusion. We also included 307 579 matched controls. Most of the study population were male (174 693 [55·7%] of the 313 837 included in the study) and of Latino ethnicity (174 906 [55·7%]). The peak association between ambient temperature and risk of acute lymphoblastic leukaemia was observed in gestational week 8, where a 5°C increase was associated with an odds ratio of 1·07 (95% CI 1·04–1·11). A slightly larger effect was seen among Latino children (OR 1·09 [95% CI 1·04–1·14]) than non-Latino White children (OR 1·05 [1·00–1·11]). The sensitivity analyses supported the results of the main analysis. Interpretation Our findings suggest an association between high ambient temperature in early pregnancy and risk of childhood acute lymphoblastic leukaemia. Further replication and investigation of mechanistic pathways might inform mitigation strategies. Funding Yale Center on Climate Change and Health, The National Center for Advancing Translational Science, National Institutes of Health.


Temperature Ascertainment
The Daymet method for calculating daily weather data in areas without weather stations relies on a combination of interpolation and extrapolation.It utilizes information from multiple instrumented locations, assigning weights to each site that consider the spatial and temporal relationships between the estimation area and the stations. 1To improve robustness, based on pre-calculated arrays of station distances, the Daymet algorithm defines a search radius for every estimation location, sized to precisely capture the average number of input stations. 1rict cross-validation analysis was performed to objectively quantify the uncertainty, with a daily mean absolute error of 1.52 °C and 1.78 °C for daily minimum and maximum temperatures, respectively. 1

Identification of Critical Windows of Exposure
The first stage association estimates were assumed to be normally distributed with mean equal to the true but unobserved association of interest, and variance equal to the squared standard error.The true associations were then modeled using the original CWVS methodology, which decomposes the single effect into continuous and binary components and uses a joint Gaussian process with temporal correlation structure to provide "smoothed" parameter estimation while conducting Bayesian variable selection output.
This second stage modeling resulted in estimates (i.e., posterior means), 95% credible intervals (i.e., equal tailed, quantile based), and relative importance estimates (i.e., marginal posterior inclusion probabilities) for each of the true associations of interest.3][4] To account for correlation between exposure windows, CWVS combines variable selection with distributed lag modeling, and uses latent Gaussian processes with time series correlation structures to model the correlation between model parameters.
From this Bayesian analysis, we calculated posterior means as the point estimates and 95% quantile-based credible intervals to quantify uncertainty in the parameters (i.e., comparable to confidence intervals in the frequentist analysis setting).For consistency with the other analyses, the credible intervals are reported as confidence intervals.
To ensure that the new methodology performed as expected, we carried out an additional sensitivity analysis where in each weekly analysis we randomly shuffled the observed mean temperature values across the cases and controls (i.e., effectively breaking any association between temperature and ALL risk).Under this randomization of exposure, we expected to not observe any statistically significant association from the second stage metaregression (i.e., credible intervals that excluded one).This is also what we observed.

Critical Window Variable Selection (CWVS) Meta-Regression
•  " !: Estimated effect for week t from the first stage regression analyses •  " !" : Variance of estimate from the first stage regression analyses (squared standard error) • : Total number of weeks in the first stage regression analyses •  !: True but unobserved effect of interest defined as •  %! : Parameters that account for temporal correlation and correlation between the continuous and binary components; modeled using autoregressive prior distributions Supplementary Figure 7. Ambient temperature and risk of childhood acute lymphoblastic leukemia among subjects with geocoding based on street address.
Legend: Sensitivity analysis of the main analysis (Figure 1 in the main text) where this analysis is restricted to subjects where street addresses were used for geocoding (1997 and onwards), as opposed to zip codes.Results from the two-stage Bayesian meta-regression analysis of ambient temperature and risk of childhood acute lymphoblastic leukemia.Accounted for race, ethnicity, birth order, maternal and paternal age, maternal education, Social Vulnerability Index, date of LMP ±7 days (i.e., seasonality and time trend), and offspring sex.Unit of exposure per 5 °C increase in mean weekly ambient temperature.Vertical bars represent 95% confidence intervals.Statistically significant positive and negative associations between ambient temperature and childhood acute lymphoblastic leukemia are highlighted in red and blue, respectively.

Gestational week
Odds ratio

Table 1 .
Number of subjects included in the main analysis, by week.The figure illustrates the number of eligible cases that were eventually included in our study.For the majority of the analyses, each case was matched to up to 50 controls (matching on sex, race, ethnicity, and date of mother's last menstrual period ±7 days), totaling 307,579 controls.In the secondary matched dataset (see main text), each case was matched to up to 4 controls (matching on sex, race, ethnicity, year of mother's last menstrual period, and residential address at birth within 10 km); here, 71 cases did not have available controls, for a final total of 6,188 cases and 24,434 controls.
Legend:Legend: Weekly means based on daily means in the warm season of cancer cases and controls matched on sex, race or ethnicity, and date of last menstrual period.