In this hospital-based study of more than 70,000 term births, we observed contrasted results according to the different air pollution metrics examined. The risk of LBW is positively associated with ambient O3 concentrations but negatively associated with NO2, as measured by monitoring stations. LBW risk is also positively associated with traffic density and proximity to major roadways, while no significant association is observed for other air pollution metrics. When birth weight is analyzed as a continuous variable, increases in mean birth weight are associated with most air pollution metrics. However, no such increase is observed for indicators of traffic density or proximity to major roadways, and a significant decrease in mean birth weight is associated with ambient O3 concentrations.
Each of the air pollution metrics that we used has respective virtues and limitations. Measurements from monitoring stations are available over large regions and effectively reflect temporal variations in the mean ambient pollution level at each monitoring site, but their lack of geographical resolution is a serious limitation for pollutants with fine scale spatial variation such as NOx and CO. Limited resolution might be less critical for pollutants characterized by strong regional components such as PM2.5. But still, the composition of particulates and their possible effects on intrauterine growth might be influenced by local sources, and thus vary at a small spatial scale
. The LUR model predictions that we used, captured small scale spatial contrasts in NO2 and NOx concentrations effectively. However, the assumption that small scale spatial contrasts in NO2 and NOx concentration remained stable throughout the study period (i.e., the assumption that at each residential address, the temporal variability of ambient concentrations was identical to that observed at the nearest monitoring site) might have been inappropriate
. Since the LUR models were developed using measurements for more recent years (2006–2007), there are uncertainties regarding their extrapolation backward in time
Measurements from monitoring stations as well as predictions from LUR models do not reflect the sole influence of local traffic as a source of air pollution. They also integrate some regional contributions from traffic and from residential emissions. Conversely, predictions from the CALINE4 model specifically capture spatial contrasts in the dispersion of primary pollutants emitted by local traffic (on roadway segments up to 3 km from maternal homes). These predictions integrate the influence of meteorology and temporal variability in emissions, though for the latter at a coarser temporal resolution than do monitoring station measurements, because of the limited temporal resolution of input data (e.g., annual average traffic flow). Since traffic flow data are also more accurate for State highways than for major roads in California because of different assessment methods
, we suspect that our CALINE4 predictions are more accurate for freeways and highways than for major roads. This would explain why CALINE4 predictions are moderately negatively correlated with distance to freeways, whereas they are weakly correlated with distance to major roads (see Table
Air pollution is a mix of a very large number of components
[9, 25, 26] and we could only measure or model a fraction of criteria pollutants which are monitored for regulatory purposes. We cannot discard the possibility that some individual air pollutants (or mix of these), which in our study setting were not strongly correlated in space or time with the criteria pollutants we focused on, might impair intrauterine growth. Considering these uncertainties, our findings based on simple indicators of traffic density and distance to major roadways are thought provoking. These indicators are arguably rather crude proxies for emission sources of traffic-related pollutants
 and may result in substantial exposure misclassification, although not necessarily in a way that would differentially affect cases and controls. Their main strengths are the integration of a spatial dimension at a very fine scale (here for traffic density, a few hundred meters or less), and their relative specificity with regard to an identified source of air pollutants. They might thus capture, though imperfectly, the effects of some primary emissions of traffic-related pollutants, or mixes of such pollutants, which were neither measured nor modeled in our study.
While traffic density within a few hundred meters from maternal homes and proximity to the nearest main road are associated with increased risk of LBW but not with changes in mean birth weight, a statistically significant decrease in mean birth weight is associated with a longer distance between maternal home and the nearest freeway. Nevertheless, this result may be sensitive to the distribution of distances. Traffic-related pollutants decay to background concentrations by 160–570 m from the edge of roadways during daytime
, and up to 2500 m downwind from freeways before sunrise
. However even under the latter condition (night time with more stagnant air), concentrations decline more sharply within 1000 m from freeways than at more remote distances
. This is likely because of the increased relative contribution of other sources (e.g. surface streets) to the pollutant concentrations at locations further away from a freeway. Therefore, the signal from freeway emissions appears to be more reasonably captured within 500 or 1000 m from freeways than at more remote distances. When analyses are restricted to subjects living no farther than 1000 m from a freeway (36% of the subjects, whereas 98% of the subjects were living within 1000 m of a main road), mean birth weight might slightly increases as distance from the nearest freeway increases, although this result is not statistically significant [see Additional file
All of our air pollution metrics share limitations that are still common in the field of air pollution epidemiology and birth outcomes
[12, 22]. The personal exposure of mothers during pregnancy, that is hypothesized to influence birth weight, could not be estimated in this large cohort since we could not take into account specific time-activity patterns
 and ambient pollutant concentrations prevailing in various living micro-environments such as the workplace or public transportation. The ignorance of these factors undoubtedly contributed to exposure measurement errors, of which direction and magnitude might differ depending on the air pollution metrics used
. This makes the comparison of results according to different air pollution metrics difficult, since each of them is probably not related to personal exposure in the same way. In spite of a few published studies on the topic
[31, 32], the relationships between personal exposure of pregnant women and the pollution metrics that we used would warrant more research in the future. In addition, our air pollution metrics relied on maternal home address at the time of delivery. For mothers who moved during pregnancy, we are unaware of the locations of other homes occupied before delivery. All these sources of exposure measurement error contribute random error to the epidemiologic results, and might also potentially generate bias.
The observed increases in mean birth weight associated with most air pollution metrics were unexpected. Since many statistical tests were applied as part of this study, with a 5% type I error risk for each, some of these significant associations might be pure chance findings. This is not likely to explain all of the significant associations we observed, however: out of the 246 statistical tests conducted, 43% were statistically significant. Although some air pollutants like diesel exhausts and polycyclic aromatic hydrocarbon have been identified as endocrine disruptor compounds
[33, 34], which are increasingly suspected to play a role in childhood and adulthood obesity
[34, 35], there is, to the best of our knowledge, no proposed biological mechanisms by which endocrine disruptors would cause increased fetal weight gain. The possible induction of gestational or preexisting diabetes by air pollution (which is a current research question
[36, 37]) might have constituted another plausible explanation for such results since both conditions may cause increased birth weight. However, our results were unaffected by adjustment for maternal diabetes. Last, a “harvest” effect of increased miscarriage among the most susceptible fetuses (which would have been more likely to have lower weights at birth) might hypothetically be triggered by exposure to air pollution and result in increased mean birth weight associated with this exposure. However, it is difficult to evaluate that hypothesis without data on miscarriages. We are not aware of any other hypothesized mechanisms for a causal relationship between exposure to air pollution and increased birth weight.
Lack of adjustment for one or several unmeasured confounders is a possible explanation for such findings. We controlled for a large set of individual potential confounders, including race/ethnicity, insurance status and maternal conditions, as well as neighborhood socioeconomic factors. Our results are robust to adjustment for further available covariates. However, we had no direct information on the height of the parents, the body mass index of mothers at the beginning of pregnancy (both usually positively associated with birth weight
) or their smoking habits (usually negatively associated with birth weight
). Adjustment for insurance status, poverty (two proxy variables for socioeconomic status) and race/ethnicity should have contributed to partially adjust for maternal smoking and pre-pregnancy body mass index. However, the possibility of residual confounding by these factors, and possibly by others (traffic-related noise, residual differences in socioeconomic status not reflected by insurance status and neighborhood-level variables, and maybe unknown factors), still remains.
If some confounding factors are truly responsible for the observed increases in mean birth weight associated with most air pollution metrics, the fact that no increase in mean birth weight is associated with local traffic density (within a few hundred meters from maternal homes) or distance to main roads might reflect different possible situations. This might indicate that the different air pollution metrics that we studied are associated in various ways with potential confounding factors. Alternatively, if all our air pollution metrics are similarly associated with potential confounding factors that bias the results “upward” (i.e., toward observed increases in birth weight), the fact that no increase in birth weight is associated with traffic density or distance to main roads might mean that they indeed exert an effect “downwards” (i.e., generate decreases in birth weight). Since these two possibilities cannot be disentangled without measurement of additional variables, the contrasting results suggests the importance of further birth weight investigations that measure air pollutants, noise, and other confounders potentially associated with local traffic.
Comparing our results to those of previous studies is not straightforward, in part because of differences in the natures and definitions of the air pollution metrics employed, varying birth outcomes of interest, as well as available information on potential confounders and strategies employed for statistical analyses. However, two recent reviews and meta-analyses attempted to summarize the findings of studies focusing on the relations between birth weight and exposure to criteria pollutants
[10, 22]. Overall, our findings of increases in mean birth weight associated with ambient air pollution concentrations measured by monitoring stations disagree with those of most published studies
[10, 11], including those conducted in California
[25, 40]. A selection effect might account for such differences. Compared to previous studies conducted in California
 and Los Angeles County
 that used birth certificates during similar study periods, our hospital-based cohort has lower percentages of Hispanic mothers (32% versus 50%
 and 70%
) and of mothers relying on public insurance (28% versus 66%
). Our cohort also has a higher percentage of Caucasian mothers (41% versus 30%
 and 13%
). The percentage of term LBW infants is also lower in our cohort (1.7%) than in the California (2.3%)
 and Los Angeles County (2.1%) studies
. Similar differences are observed when our cohort is directly compared to birth certificates from Los Angeles and Orange counties for year 2001 (that is, in the middle of our study period) [Additional file
In a recent meta-analysis by Stieb et al.
, ambient ozone concentrations were associated with a non-significant reduction in mean birth weight. Previous studies conducted in California reported significant decreases in mean birth weight associated with ozone concentrations
[40, 41], but a recent one found no association with LBW
. Results from the latest studies conducted in other settings are also mixed, ranging from significant decreases in birth weight
 or increases in the risk of being small for gestational age
 to reduced risk of LBW
A rapidly growing number of studies have used predictions from LUR models for NOx, NO or NO2, with some results indicative of decreases in birth weight (e.g.,
[9, 25]), while others are not significant (or mixed depending on the birth outcome considered) (e.g.,
), or indicate an apparent increase in birth weight
. Only one used a line source dispersion model specific of traffic emissions, comparable to CALINE4
. It reported a birth weight reduction associated with CO, in a selected sub-area of the study setting
. Traffic density indexes, with varying definitions, have been used in seven studies. Four of these reported decreases in birth weight
[27, 45–47], two reported null or mixed results
[37, 48] and one an apparent increase in birth weight
. Proximity to roadways was associated with decreases in birth weight in six studies
[30, 37, 46, 49–51], while two others reported no significant associations