Preterm birth is a major public health burden whose prevalence continues to rise. The rate of preterm birth in the U.S. is 12.7%, the highest rate reported to date[1]. The financial burden of caring for infant survivors of preterm birth is substantial, with a lower-end estimate of approximately 26.2 billion dollars annually in the U.S.[2]. Preterm birth is a complex disease with etiologic influences from a variety of social, environmental, hormonal, genetic, and other factors[3]. Our limited understanding of the complex interactions among these factors contributes the lack of effective intervention strategies available to reduce the rate of preterm birth.

It has been hypothesized that the association between preterm parturition and individual-level socioeconomic deprivation is primarily accounted for by the co-existence of other significant risk factors such as medical comorbidities, lack of prenatal care, and adverse behaviors (smoking and alcohol use) which occur more commonly in women of lower socioeconomic status[4]. However, many prior studies have been limited by design constraints and the inability to thoroughly account for potential confounding factors.

Implementation of appropriate statistical methodologies, rigorous adherance to study design, and consideration of as many clinically important covariates as possible optimizes the capability to accurately quantify the independent association between area-level poverty and preterm birth risk. When taking into consideration that people who reside within the same area share common area-level neighborhood conditions, individual-based statistical models such as traditional logistic regression analyses are not optimal to determine the true strength of the association. Using multilevel logistic regression analysis has the advantage of allowing one to estimate not only the fixed effects of individual-level and area-level factors, but also the random effect of geographic variation of preterm birth between areas.

The purpose of this study was to utilize a large population-based birth registry to estimate the independent effect of county-level poverty on preterm birth risk. To accomplish this, we used a multilevel logistic regression approach to account for multiple co-existent individual-level variables and county-level poverty rate.

A data set which included Missouri birth records from 1978–1998 was provided in a de-identified fashion for this analysis by the Missouri Department of Health and Senior Services. The study was considered exempt from review by the Missouri Department of Health and Senior Services IRB and the Human Subjects Committee of Washington University in St. Louis. This data set has been a rich source for the analysis of factors associated with birth timing [5–10].

A total of 1,577,082 births occurred in Missouri between 1978 and 1998. To optimize our capability to examine the influence of important individual factors, we limited our study to those births which occurred between 1989 and 1998 due to an unacceptable amount of missing demographic data for births that occurred before 1989 due to revisions in the birth certificate format in 1989 and improvements in edit and query systems after that date. Because deliveries which occur prior to 20 completed weeks of pregnancy are considered miscarriages rather than births, we limited our analysis to births recorded as having occurred at 20 weeks or greater. We excluded births resulting from a multifetal gestation, intrauterine fetal demise, or involving a major congenital malformation due to their know propensity to deliver preterm, potentially for mechanisms unrelated to the exposure we wished to evaluate. After these exclusions, the study population consisted of 675,044 births. We further limited our analysis to births occuring to mothers whose reported residence at the time of delivery was in the state of Missouri. There were 40,050 births in Missouri to mothers who resided in other states during the study period, yeilding a final population of 634,994 births available for analysis.

In this study we found that women who gave birth in counties with the highest poverty rates were at increased risk of preterm birth at < 35 and < 32 weeks of gestation. Although these risks are modest, with odds ratio estimates of 1.18 and 1.27 respectively, the confidence intervals consistently excluded the value of one even after accounting for numerous individual demographic, obstetric, behavioral, and medical risk factors. These effects were similar regardless of maternal race. Modest increases in MOR and IOR values suggest that county-level poverty was an important area-level characteristic for geographic heterogeneity of PTB. But, other unmeasured area-level characteristics may also partly contribute to the geographic heterogeneity of preterm birth.

Other investigators have reported an association of individual and area-level socioeconomic deprivation with adverse perinatal outcomes, such as low birth weight and preterm birth[5, 17–26]. Although numerous, these studies vary significantly with respect to sample size, study design, and methodologic strategies. Although women who are economically disadvantaged are disproportionately likely have co-existent risk factors for preterm birth, attempts to adjust for these factors in prior studies have not completely eliminated the modest association between poverty and prematurity. Our goal was to use optimal statistical techniques on a large-population based sample to describe the effect of regional poverty on prematurity. Due to the variety of information available, we were able to evaluate the effect of several partitions of clinically relevent social and medical factors on preterm birth while accounting for the random variation of prematurity across counties. Despite accounting for many sociodemographic, obstetric, medical, and other factors our results suggest that county environments may in fact independently influence birth outcomes.

The advantage of multilevel analytic approaches to health outcome research is that they enable us to evaluate and quantify the heterogeneity of the outcome across geographic areas. When minimal geographic heterogeneity exists, this implies high area-level correlation. When individuals within the population are similar and vary little with respect to health, this provides optimal conditions for regional prevention strategies[27]. Likewise, health outcomes such as preterm birth, which can be significantly influenced by behavior-related factors (smoking, alcohol use, receipt of prenatal care) may be less influenced by area-level factors than diseases with a long natural history such as atherosclerotic disorders. Also, considering that area-level effects are minimized as individuals move geographically,[15] it is unlikely that population mobility significantly accounted for our findings given the finite time of the gestational period. In the case of preterm birth, a serious complex disease process whose incidence continues to rise, refining our ability to identify at-risk populations in which to target preventative health care strategies has particular appeal.

The datasource used for this study has many advantages such as its large size and population-based nature. It includes a breadth of demographic information regarding the parents, detailed medical and obstetrical characteristics of the mother's antepartum and intrapartum course, as well as specific information regarding the status of the neonate at birth. Despite the extensive data available for analysis regarding each birth, there are some limitations to consider as well. One of the most commonly cited limitations of birth outcome research from vital statistics data is validity of the recorded gestational age[28]. In an effort to minimize the effect of gestational age estimation inaccuracies and potential misclassification bias,[12] we chose our primary outcome as birth prior to 35 weeks of gestation rather than the traditional definition of preterm birth, less than 37 weeks. The data set also had no information which could be used to reliably delineate spontaneous from medically indicated preterm births. Other limitations inherent to the datasource are more difficult to adjust such as the potential for undercoding or underreporting of obstetric and medical risk factors and errors in birth certificate data extraction and entry. A large amount of underreporting of important risk factors, if non-random, could have resulted in an increased effect size estimation of poverty on preterm birth risk due to the inability to adjust for the unreported risks in multivariable analyses. It is unlikely, though, that significant underreporting of the most important obstetric and medical risk factors occurred within the database period we evaluated as the relative prevalence of those included in our analyses are consistent with known rates elsewhere in the US[9]. On the other hand, many social and demographic variables are likely to be accurate, such as race, level of education, receipt of state-funded assistance programs, and county of residence.

Another limitation of the study is the lack of ability to account for other county-level factors which could have possibly influenced the association between poverty and preterm birth risk, such as regional crime rates, adverse environmental exposures, etc. We incorporated as many clinically importanty concomitant factors into our analysis as were available through the data source, but we acknowledge that other unmeasured and unknown potential co-existant risk factors could potentially exist.

In this study we chose counties as the geographic units of analysis. There are several reasons for this approach. US counties are the smallest geographical entity within a state with the social, political and legal responsibility for providing a broad range of social services and health care. Counties may qualify as communities to the extent that individuals and groups within them participate in community development by identifying county-wide problems, collecting health, social and environmental data, and by formulating and implementing specific public policy measures[29]. Counties are also the smallest geographical entity for which health, socioeconomic, and population statistics are consistently available over time. They are stable sociopolitical and geographic entities, but also provide an apppropriate socioeconomic, political, and community context within which many social and public health policies are formulated [28, 30–32]. From a governmental perspective, using a county-based approach, rather than examining outcomes by zip code or census tracts, allows for organized implementation of healthcare policies and interventions.

In conclusion, county-level poverty was associated with an increased risk of preterm birth. Although the association is modest, even a modest increase in preterm birth risk as demonstrated in this study has the potential to have a robust effect given the large number of women residing in areas with relatively high poverty levels.

Individual-level and area-level poverty have been associated in the past with an increased risk of preterm birth, but prior studies have been limited by design constraints and the inability to thoroughly account for potential confounding risks. We utilized a large population-based birth registry from the state of Missouri to perform a multi-level analysis of the association between poverty and preterm birth while accounting for many potential co-existent risks via optimal methodologic strategies. The findings from our study provide obstetrical care providers and health-care policy makers with important information regarding the prevalence of preterm birth in counties with high poverty and an accurate estimate of the effect of high area-level poverty on preterm birth risk.