The incidence of low and very low birthweight — i.e., birthweights below 2500 and 1500 grams, respectively — are important indicators of public health. Not only is low birthweight a leading risk factor for infant mortality, it has also been linked to an increased risk of serious long-term disabilities (Hack et al., 1995) and onset of chronic diseases in adulthood (Collins and David, 2009, Barker et al., 1995, Rich-Edwards et al., 1999). Worldwide, it is estimated that 15%–20% of all births — and more than 20 million births per year — have low birthweight (World Health Organization, 2014). While the incidence of low birthweight in the United States (U.S.) is much lower, 2017 marked the highest level in over 20 years (8.27%) — a value that was subsequently exceeded in both 2018 and 2019 (Martin et al., 2021). Of course, it is also true that the incidence of low birthweight is not equal for all populations. Reports from the Centers for Disease Control and Prevention (CDC) indicate increasing racial disparities in low birthweight from 2015 to 2019, with the incidence of low birthweight for Black mothers more than twice that of White mothers (Martin et al., 2021). Furthermore, geographic disparities in the incidence of low birthweight are also observed, with rates in Mississippi (12.3%) being nearly double those in Alaska (6.3%) (Martin et al., 2021).

While the topics of racial and geographic disparities are important in their own right, it is also important to study geographic trends in racial disparities, as this can shed light on regions in which all race/ethnicities have similar health outcomes and regions where some racial/ethnic populations are being left behind. For instance, Goldfarb et al. (2018) assessed the incidence of low birthweight in 400 of the largest U.S. counties from 2003–2013 and found that only four counties demonstrated improvement toward racial equality via improvement in the incidence of low birthweight for Black mothers (rather than worsening trends for White mothers). Unfortunately, such studies of low birthweight are plagued by two key issues. First and foremost, while the CDC’s Wide-ranging Online Data for Epidemiologic Research (WONDER) is a great resource for accessing mortality data, data available from CDC WONDER pertaining to births is limited to counties with population sizes larger than 100,000. Not only does this represent fewer than 20% of U.S. counties, but it also precludes inference on urban/rural disparities in the incidence of low birthweight. The second key challenge encountered when investigating geographic trends in racial disparities in low birthweight is that even when access to detailed county-level data is available via state-specific web portals — e.g., the Commonwealth of Pennsylvania’s Enterprise Data Dissemination Informatics Exchange (EDDIE) system — many parts of the country have small racial/ethnic minority populations. Because counties in these areas are more likely to have few births to non-White mothers, estimates of the incidence rates of low birthweight births for minority populations may be unreliable or otherwise lack the level of precision required for inference.

The objective of this paper is to develop a modeling framework that offers a compromise between increasing the precision of estimates of rates such as the incidence of low birthweight and protecting against oversmoothing — i.e., when the contribution of the model overwhelms that of the data and distorts the underlying trends. As a starting point, we consider the use of Bayesian methods that incorporate spatial structure into the model, such as the conditional autoregressive (CAR) model framework of Besag et al. (1991). Aside from its use in a variety of applications (Waller et al., 1997, Gelfand and Vounatsou, 2003, Quick et al., 2015, Botella-Rocamora et al., 2015), recent work by Quick et al. (2021) has illustrated the extent to which the CAR model framework can oversmooth estimates by virtue of putting excessive weight on the model’s spatial structure, thereby overwhelming the contribution of the data from individual regions. For instance, Quick et al. (2021) analyzed county-level, heart disease-related death data and estimated the contribution of the model as being equivalent to 36 additional deaths occurring in each county, despite many counties having fewer than 10 observed deaths.

In this paper, we extend the work of Quick et al. (2021) from the Poisson-distributed count setting to the case of binomially distributed data for the purpose of investigating racial disparities in the incidence of low birthweight in Pennsylvania counties. In Section 2, we describe the data used in this analysis. Section 3 provides a brief overview of the approach of Quick et al. (2021) for quantifying (and restricting) the informativeness of the Besag et al. CAR framework and describes its extension to binomial data. We then evaluate the properties of our approach via simulation in Section 4 prior to implementing our approach to investigate trends in low birthweight in Section 5. We then conclude with a discussion in Section 6.