Some of the earliest research on neighbourhoods and health is in perinatal epidemiology. For example, in 1947, Yankauer presented research at the American Public Health Association Annual Meeting showing neighbourhoods in New York City with a high proportion of Black births had a higher infant mortality rate among both Black and White births 1. In this issue of Paediatric and Perinatal Epidemiology, Dixon and colleagues 2 build on this long-standing body of research by seeking to provide evidence for downstream interventions that can buffer upstream neighbourhood harms for pregnant people. This study stimulates discussion of investigating individual-level pregnancy interventions to buffer neighbourhood harms. The goal of this commentary is to contribute to this discussion by focusing on several challenges to causal inference and on suggestions for future directions in perinatal epidemiologic research. Dixon et al. tested whether social support buffers the association between area deprivation (ADI) and both preterm birth (PTB) and small-for-gestational-age (SGA) birth using data from NuMoms2b, a pregnancy cohort of nulliparous women across eight U.S. clinical research sites. They first examined the main effects of both ADI and social support on PTB and SGA, then tested additive interaction between ADI and social support, followed by three-way additive interaction with race/ethnicity. Social support is an important, more modifiable downstream factor than ADI, and a key component of many pregnancy interventions such as doula care, group prenatal care, and community health workers. Its potential to prevent adverse birth outcomes shows promise for the power of community and for what tangible support for mothers could look like. Some results supported their interaction hypotheses—women with high social support had a lower risk of PTB than those with low or moderate support, and high social support buffered the association between ADI and SGA in white women. However, other findings were unintuitive, for example, the harmful effects of ADI on PTB seemed mostly limited to those with high social support. What does this tell us about challenges in studying neighbourhood factors, birth outcomes, and potential interventions at the individual level? Causal inference challenges in neighbourhood effects research, including in perinatal epidemiology, are well-established 3, but are less often considered in the context of cross-level interactions between neighbourhood exposures and individual-level interventions. Central to the issue of achieving exchangeability between cohorts exposed and unexposed to ADI is the issue that pregnant people may be selected into neighbourhoods based on individual factors, which are causes of adverse birth outcomes. Examining individual buffers of neighbourhood harms by estimating interaction requires conditional exchangeability: pregnant people in deprived versus non-deprived neighbourhoods must be comparable with respect to causes of adverse birth outcomes and also comparable within strata of social support (and, for three-way interactions, within strata of race/ethnicity). Moreover, if this cross-level interaction plays out over time, for example, the ADI exposure precedes the social support intervention, time-varying confounding may additionally complicate the goal of conditional exchangeability. Dixon et al. appropriately did not adjust for individual-level risk factors for PTB, such as pre-pregnancy hypertension, that may lie on the causal pathway from ADI to PTB; however, because chronic hypertension precedes the hypothetical social support intervention, it may confound comparisons between social support groups. Therefore, individual health conditions, such as pre-existing hypertension, may be mediators of the exposure and outcome but are post-exposure confounders of the intervention and outcome. To address this, the use of g-methods is a preferred alternate 4. Issues of selection into neighbourhoods also connect to a second identification challenge: positivity. Historical forces such as racial housing segregation and economic discrimination create issues with structural confounding, such that it is often impossible to model individual selection into neighbourhoods (to achieve exchangeability) without running into positivity violations between deprived and non-deprived neighbourhoods 5. The assumption of positivity means that every individual must have a non-zero probability of living in a deprived or non-deprived neighbourhood. Sparse cell sizes, whether due to sample size limitations or structural violations (not only in the study sample but also in the target population, where no individuals exist in some strata), can threaten exchangeability. These problems may contribute to some of the non-intuitive findings of Dixon et al. Some study designs to address these challenges include comparing births from the same mother who moved between neighbourhoods, restricting the analysis to neighbourhoods with a similar propensity for the exposure, comparing changes within neighbourhoods, and quasi-experimental designs that assess the impact of neighbourhood improvements on birth outcomes. Other relevant causal inference challenges for neighbourhood and health research, such as consistency and interference, have been discussed at length previously 6, 7. These challenges also apply to studies of the interaction between neighbourhood exposures and individual interventions. For example, the apparent buffering of ADI in high-support neighbourhoods could partly reflect different conceptualizations of the construct ‘social support’ (a violation of consistency), rather than true interaction (which could explain the finding that high social support buffered ADI only among white women). The dispersed geographic nature of NuMoms2b may help protect against interference, though within study sites, one pregnant person's social support may have affected others, resulting in partial interference 8. Issues of causal inference highlight the strengths and weaknesses of NuMoms2b and other clinical pregnancy cohorts for testing hypotheses about potential interventions to buffer neighbourhood harms. The main strength of a clinical pregnancy cohort is the wealth of information from questionnaires, biological specimens and prospectively collected medical records. These features contrast with the limitations of birth records or claims data, which have little information on psychosocial factors, rarely have biomarkers and are often cross-sectional. Researchers could lean on these strengths in several ways. A wide range of covariates could be used to derive the inverse probability weights to account for neighbourhood selection to improve exchangeability. Information on miscarriages and stillbirths could be used in the analytic study design to reduce live birth bias. For example, if ADI is a cause of both foetal loss and preterm birth, and unmeasured factors are also causes of both, then the risk of PTB in high ADI neighbourhoods could be biased downward due to higher foetal losses in this group 9. Live birth bias could also affect associations with SGA, which is further complicated by known limitations to the use of SGA as a proxy of foetal growth restriction 10. Finally, a major advantage of clinical pregnancy cohorts is the greater validity of clinical outcomes and the ability to more precisely define them than in administrative data. NuMoms2b has information on indications for preterm birth that could be leveraged to refine hypotheses—for example, does social support buffer from the stress of high ADI neighbourhoods, and therefore have the most impact on spontaneous preterm birth or early preterm birth? Or does it increase access to high-quality health care, and therefore reduce medically indicated preterm birth? Despite the inherent challenges to studying upstream and downstream causes simultaneously, future research could benefit from this ‘multi-stream’ thinking. Upstream thinking regarding the structural determinants of poor birth outcomes encourages researchers to consider the pregnant person's neighbourhood environment when evaluating potential interventions such as social support. Conversely, tests of neighbourhood interventions could leverage individual data to examine differential impact based on individual characteristics; for example, adding a new park to a neighbourhood may only increase physical activity for those with a supportive walking partner. Evaluating potential mediators of upstream causes could help identify mechanisms to reduce disparities, but mediation analyses often require stringent assumptions that can be difficult to justify and may require the use of g-methods to address time-dependent confounding. Refining approaches to studying multi-stream causes in perinatal epidemiology can improve our understanding of adverse pregnancy outcomes and identify interventions to improve maternal and infant health. The authors have nothing to report. The authors have nothing to report. The authors declare no conflicts of interest. This is a linked article to Dixon et al. ‘The Role of Social Support as a Buffer Against Adverse Birth Outcomes Among People Experiencing Neighbourhood Deprivation.’ To view this article, visit https://doi.org/10.1111/ppe.70113. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.
Janevic et al. (Mon,) studied this question.