Live Birth Bias May Play a Role in Epidemiological Analyses of Air Pollution and Autism Spectrum Disorders

Background: There has been increasing interest in the scientific literature recently on the role of air pollution exposures on the development of autism spectrum disorder (ASD). Findings and conclusions across epidemiological studies, reviews and meta-analysis however, are not consistent. In a recent publication testing the associations between prenatal exposure to NO₂ and autism, a distributed lag model was implemented with population-based ASD data and weekly NO₂ exposures from Israel in order to identify critical windows of vulnerability. When mutually adjusted in a distributed lag model, postnatal exposures presented positive associations (more NO₂, more ASD) while prenatal associations varied by week of pregnancy, and gradually reached a negative, statistically significant peak around the end of the first trimester. We suggest that the negative associations do not represent protective effects of air pollution, but are the result of live birth bias.

Objectives: Explaining the causal structure and the assumptions needed for live birth bias to create biased negative associations between air pollution and ASD.

Methods: A directed acyclic graph (DAG) was built to represent the causal structure and the underlying assumptions.

Results: As presented in our DAG (see graphical abstract), live-birth bias could arise from the fact that ASD can only be assessed in live-born children, and many pregnancies do not end in a live birth. This inevitable selection of only live births into the analysis in question may lead to bias of the observed association from the actual causal association if a) air pollution is a risk factor for pregnancy loss, and b) there are other factors ("U", likely unmeasured, even unknown) that influence both pregnancy loss and ASD. In the causal inference terminology, the selective analysis of only live births opens the backdoor path: ASD <-- U --> Pregnancy Loss <-- Air Pollution, which creates an association between air pollution and ASD and biases the causal association in question, which is represented by the dashed arrow between Air Pollution and ASD.

Conclusions: We suggest that live-birth bias can create an observed negative association between air pollution and ASD. Several lines of evidence in the literature support the first assumption of air pollution as a risk factor for pregnancy loss. The second assumption is harder to assess, since U is undefined, but one possible example is prenatal stress: the existing literature supports its involvement as a risk factor for both pregnancy loss and ASD in the offspring. Thus, prenatal stress may be a possible example for our variable "U", although the suggested bias mechanism is not limited to this specific factor. This bias has implications for all air pollution-ASD studies, and it may also be relevant to other neurodevelopmental conditions. It could create an apparent protective association, and it could mask an increased risk.