In 2016, the proportion of births delivered preterm (defined as spontaneous or iatrogenic delivery before 37 weeks of completed gestation) in the United States was 9.85% [1]. Preterm birth (PTB) is a public health concern because it is associated with high infant mortality and adverse developmental outcomes later in childhood [2]. Known risk factors for PTB include low socioeconomic factors, age, race, substance abuse, tobacco usage during pregnancy, poor nutritional status, and the presence of a birth defect [3,4]. Recently, some epidemiologic studies have suggested that exposure during pregnancy to ambient particulate matter less than 2.5 μg/m³ in aerodynamic diameter (PM_2.5) is associated with an increased risk of PTB; [5–8] however, results have been inconsistent, and the exact biological mechanisms surrounding that relationship and any potential critical exposure windows remain unclear. Ambient outdoor PM_2.5 is a ubiquitous pollutant for which personal exposure may vary on a number of factors, including proximity to emission sources (e.g., mobile vehicles) [9].

Whereas studies have generally shown an increase in PTB with an increase in PM_2.5 exposure [10], studies on specific, and usually narrower, windows of exposure have been less consistent. Identifying these critical exposure windows may aid in determining the exact biological mechanism(s) behind the relationship. Some studies have provided support for early exposure effects of PM_2.5 on PTB (i.e., within the first trimester) [5,8,11,12]. Conversely, others have suggested that exposure during the entire pregnancy and late pregnancy (i.e., during the third trimester), but not during the first trimester, has a stronger association [13], and additional research has reported inverse or null results [14,15]. In addition, there has been a growing body of literature surrounding the relationship between PTB and ambient temperature [16,17], with higher temperatures being positively associated with PTB. For example, study by Schifano et al [18] evaluated PTB’s relationship with PM₁₀ stratified by season and found that temperature impacted the relationship in the warm season, but not in the cold season. Conversely, in a meta-analysis of the effects of seasonality on the PM_2.5 and PTB relationship, authors reported that both summer and winter seasons (compared to the more temperate autumn and spring) may act as modifiers [19]. Although there is limited evidence looking at specific biological mechanisms that may explain this relationship, multiple studies have shown that exposure to PM_2.5 can cause placental inflammation during pregnancy [20] potentially resulting in altered placental vascular function [21], which in turn may cause inadequate nutrient exchange or placental perfusion [22,23] leading to PTB. While the ways temperature may act as an effect measure modifier on the PM_2.5 and PTB relationship remain unclear, limited toxicological evidence suggests that exposure to concurrent high temperature may intensify the effects of an environmental exposure [24]. Narrowing down when this relationship is the strongest could aid in identifying more specific biological mechanisms behind the relationship and could allow for targeted interventions.

Given the inconsistent relationship between PTB and PM_2.5 and the potential for high temperature to have an impact both directly on PTB as well as on the intensity of an effect from an environmental pollutant, our study was designed to investigate the relationship between PM_2.5 exposure and PTB during various potential critical windows of exposure and then to further investigate whether the potential relationship is modified by concurrent exposure to high temperatures. Specifically, we hypothesize that exposure to PM_2.5 during the first few months of pregnancy could interfere with early placental development (likely by causing placental inflammation during early pregnancy), which in turn could result in PTB. We further hypothesize that this relationship would be stronger with the added stress of higher ambient temperatures. To evaluate these hypotheses, we conducted a case-control analysis using the controls who participated in the National Birth Defects Prevention Study (NBDPS).

Overall, 2839 births were included in these analyses, 246 PTBs as cases and 2593 term births as controls, though the number of births in each model varied slightly. In general, the study population was composed primarily of non-Hispanic white women with some postsecondary education, household income between $10,000 to $50,000 annually, born in the United States, and non-smokers (Table 1). The average interquartile range for PM_2.5 concentrations across states was 4.7 μg/m³, and temperature distributions varied between states (Table 2). There was no appreciable or consistent difference between the PM_2.5 concentrations that were stratified and not stratified by high average temperature, suggesting that in stratifying high temperature days, we were not inadvertently also stratifying by high PM_2.5 days. In addition, quartiles of temperature did not vary substantially between cases and controls (Table A3). Most preterm births were less than 37 weeks of gestation and greater than or equal to 32 weeks (n = 222), with few births occurring less than 32 weeks (n = 24). The average distance between the first reported maternal residence and the closest monitor for the first eight weeks of pregnancy was 13.3 km, with a median distance of 9.2 km (SD: 11.4 km).

When the PM_2.5 and PTB relationship was modeled using categories of PM_2.5 exposure, we observed a positive relationship between PTB and PM_2.5 for the third and fourth months of pregnancy (Table 3). We used the full models for all subsequent analyses (i.e., temperature stratified and weekly PM_2.5 exposure). We did not observe the relationship that we saw in the categorical models in months 3 and 4 in any of the continuous models; however, that could be due to potential departure from linearity not captured in the continuous model rather than a true lack of association, though we note that we did not test for departures from linearity, rather we opted to reported categorical exposure results.

In addition, we conducted analyses for months 2 through 4 of pregnancy in which PM_2.5 concentrations and temperature were averaged over each week instead of each month, to see if there were specific weeks where the associations were stronger (Fig. 1 and Table A1). We observed that weeks 9 through 12 all had point estimates that were elevated, corresponding to months 3 and 4, respectively; however, there was no single week that stood out as demonstrating a stronger association between PM_2.5 exposure and PTB.

In the models where we assessed effect measure modification by average temperature during pregnancy, we found that in almost every instance, the low temperature group, representing the state-specific lower 75th percentile of temperature during pregnancy, showed no relationship between PTB and PM_2.5. These models were adjusted for local temperature and all other covariates in the nonstratified full models. In addition, the low temperature group displayed consistently lower point estimates than the upper 25th percentile group for the months and surrounding months where we saw a relationship in the nonstratified results (Fig. 2 and Table A2), though we note that credible intervals overlap. Although we observed a positive relationship between PM_2.5 concentrations and PTB in months 3 and 4 in the nonstratified analyses, in the analysis stratifying by temperature, the effect of PM_2.5 in months 2 through 5 was elevated for the high temperature group. Although we cannot directly compare the two strata, we can compare each to the crude estimates and note that there is a trend of the upper 25th percentile being higher than the crude estimate, while the lower 75th percentile is generally below it; however, results were imprecise and credible intervals overlapped (Fig. 2). In addition, the length of pregnancy for the cases in the high temperature group was 2.4 days shorter than the length of pregnancy in the cases in the low temperature group.

The elevated odds ratios (ORs) that we observed for exposure in months 3 and 4 are consistent with recent studies that reported elevated risk associated with exposure averaged over the first [37,38] or second trimester [37,39]. In our study, the months we found to have elevated effect estimates would generally be split into two different exposure periods in analyses that average exposure over trimesters (i.e., exposures during month 3 would be assigned to trimester 1 and exposures during month 4 would be assigned to trimester 2). If PM_2.5 exposures during months 3 and 4 are etiologically relevant to PTB, evaluation of exposure averaged over trimesters could explain some of the inconsistencies observed in these studies.

When we evaluated individual weeks of exposure, we observed elevated ORs during weeks 9–12, though no single week stood out as being etiologically important. We note that concentrations between weeks tended to be correlated, thus limiting our ability to detect critical periods of exposure, should they exist. Several other recent studies have conducted analyses of weekly exposure periods. Using a time-series design by Arroyo et al [40], observed positive associations with exposure during week 17 of gestation. [41] and [11] separated PTB into multiple categories based on gestational age and both observed positive and negative associations depending on combined exposure and outcome period [41], with 4-week exposures, and by Rappazzo et al [11] with exposures during individual weeks of pregnancy [42]. examined the relationship between PTB and traffic-related air pollutants, including PM_2.5 over each trimester, the entire pregnancy, and over the last six weeks of pregnancy and generally observed a positive relationship between PM_2.5 and PTB in the second trimesterand over the entire pregnancy regardless of PTB subtype. Generally, the results of weekly PM_2.5 exposure analyses and PTB have been inconsistent across studies, and no weekly (or subweekly) period of PM_2.5 has been identified as etiologically relevant for PTB. Although we recognize that controlling for other exposure windows (e.g., other months of pregnancy) would be preferable, we lacked the sample size to adequately do so. In addition, there is also evidence from experimental and epidemiologic studies demonstrating that PM_2.5 exposure could result in a coherent series of physiological responses that provide biological plausibility for PTB, including placental oxidative stress and intrauterine inflammation, altered fetal metabolism, altered placental growth, and impaired implantation [43].

Our study is novel in its examination of ambient temperature as a potential effect measure modifier of the relationship between PM_2.5 exposure and PTB. Recent studies have evaluated the effect of ambient temperature on risk of PTB and have reported greater risks of preterm birth in summer and winter (when temperatures are most extreme) compared to spring and autumn (when temperatures are more moderate) [19]. We are unaware of any studies that have evaluated how temperature could modify the relationship between PM_2.5 and PTB risk, though by Schifano et al [18] looked at the relationship between PM₁₀ and PTB stratified by season, which is likely a proxy for temperature, and found a potential effect of heat and PM₁₀ exposure on PTB risk. Our results, albeit imprecise, demonstrate the ability of higher temperatures to modify the effect of PM_2.5 exposure modeled as a continuous variable on risk of PTB. The potential biological mechanisms behind this relationship are unclear; however, limited toxicological evidence suggests that exposure to high ambient temperatures may intensify the effects of environmental pollutants, thus amplifying a relationship between PM_2.5 and PTB [24]. In addition, physiologic changes in response to higher temperatures, such as increased ventilation, can play a role in increasing the biological dose of an air pollutant that is experienced among exposed populations, so concurrent high temperature and higher PM_2.5 may result in a higher dose of PM_2.5 than the mother would have under lower ambient temperatures. Finally, atmospheric processes could mitigate direct effects of temperature on air pollution [44]. For example, temperature inversions are capable of trapping air pollutants, including PM_2.5, closer to the ground and increase the potential for exposure [45].

This study has several strengths, including the geographic scope of the NBDPS and the availability of complete residential histories, which aids in reducing exposure misclassification due to residential mobility and relying solely on residence at the time of delivery to assign exposure [46,47]. Centralized geocoding increased the consistency of the data across the NBDPS participating centers and improved the quality control of the geocoded data used to assign exposure. Studying NBDPS controls allowed us to investigate PTB without accompanying birth defects. This is an important consideration as certain birth defects have been observed to be higher among PTBs [48] and excluding infants with birth defects could be removing an important potential confounder, operating under the hypothesis that infants with birth defects are not a mediator on the PTB and PM_2.5 relationship. In addition, the NBDPS controls have been demonstrated to be representative of their source population [27], improving the generalizability of our results.

Our study also had limitations. Specifically, our study was not able to account for different subsets of PTB, which may have different predominant causes. In addition, our study did not have information on spontaneous compared to iatrogenic PTB; however, we hypothesize only spontaneous PTB would be associated with exposure to PM_2.5. Both of these limitations would likely lead to bias toward the null. Our data were also limited by sample size, and a relatively small number of cases overall may have precluded us from seeing a relationship in certain analyses, though we note that it is likely that we would only see effects during biologically relevant periods of exposure, which are unlikely to occur during the entire length of pregnancy. In addition, as in many PTB studies, our study used estimates of gestational age to assign case status, and inaccuracies in predicting gestational age may make it more difficult to identify critical windows of exposure, particularly in our weekly analyses.

Another potential limitation is the inability to distinguish between the potential effects of temperature versus ozone in both the temperature stratified and nonstratified analyses. Although our study both controlled for and stratified by high temperature, it is possible that ozone concentrations may actually be driving the effect measure modification we saw, as ozone and temperature are generally positively associated [49]. To our knowledge, there are no studies that have examined ozone as a potential effect measure modifier in the PM_2.5 and PTB relationship, and studies that have looked at the ozone and preterm birth relationship directly show mixed results [6,50,51].

With regard to our temperature-stratified analyses, we only stratified by high temperature; however, if very low temperature days also acts as an effect measure modifier on the PTB and PM_2.5 relationship, then it is possible that in certain cases (i.e., during very cold winters), high temperature could be protective in the winter. If this is the case, we would expect to see our high temperature strata biased toward the null, as both summer and winter months are included in the analysis.

An additional limitation is our assumption of independence for our models, as we recognize that the preceding month and proceeding months of interest are likely associated with the month analyzed (e.g., the exposure in month 3 is likely correlated with the exposure in months 2 and 4). This limits our ability to draw conclusions about any individual month as a critical window of exposure. Nevertheless, although we are not able to fully parse out that a single month of exposure is more strongly associated with PTB, we do see a pattern of increasing ORs during specific adjacent months, suggesting that a critical window, should one exist, may be contained in those months.

To ensure that we captured the largest amount of study population possible, while still using monitors that could presumably capture or be correlated with a person’s true exposure, we used monitors up to 50 km away; however, this large distance has the potential for increased exposure measurement error, particularly if the monitors used are not correlated with the true exposure concentrations. Previous work has demonstrated that risk estimates based on the nearest monitor may be more biased than those based on land use regression models or kriging [52]. Despite choosing a 50 km monitor distance, a portion of participants had to be excluded because of missing exposure data (n = 2872). We anticipate that this maylead to bias toward the null; however, if the lack of monitor data within 50 km was systematically due to something that was related to both exposure and outcome (e.g., if the excluded participants consistently had lower exposures to PM_2.5 during relevant time periods and higher rates of PTB), it is possible that the bias could be away from the null. We further note that although we tried to use the largest study population possible, a portion of the participants had to be excluded because of missing covariate data (n = 629), which has the potential to bias results if important covariates differ substantially between the included and excluded subjects. Finally, if mothers mediate their exposure to stay inside, then ambient, outdoor PM_2.5, and temperature concentrations may not be associated with their true exposure, leading to bias toward the null.

In addition to spatial scale, we recognize that the temporal scales we used (i.e., months and weeks) may bias results, particularly given that we do not yet have a firm understanding of the critical exposure window [53]. Nevertheless, we tried to choose temporal averaging windows that we thought would be the most sensitive to our outcome based on the toxicological and epidemiologic information currently available, and less sensitive to variations in PM_2.5. Thus, in our main analysis, instead of focusing on the trimester, which we anticipated may mask an effect or the week, we focused instead on the month. We expect this bias related to our temporal window choices to be nondifferential with regard to outcome status.

Overall, we observed elevated risks of PTB associated with PM_2.5 exposures during months 3 and 4 of pregnancy. These risks became more apparent when the exposure period coincided with periods of higher ambient temperatures in stratified analyses, suggesting that temperature may modify the relationship between PTB and PM_2.5, though we note that results were imprecise and any interpretations as to the true relationship between these covariates must be made with caution.