Preterm birth, as defined by the World Health Organization,11 is a delivery which occurs at less than 37 completed weeks of gestation. We performed our primary analysis with preterm birth defined as less than 35 completed weeks in order to enrich for a population of deliveries which were truly preterm by avoiding births occuring at borderline gestational ages between 35 and 37 weeks, in an effort to minimize misclassification bias12. We defined early preterm birth as birth occuring prior to 32 completed weeks of gestation, because the risk of neonatal morbidity is higher for births of shorter gestations.

We included the following individual-level measures: maternal age, maternal race (self-reported), maternal and paternal highest educational attainment, residence within city limits, birth sequence, marital status, presence of medical risk factors, marital status, indicators of low income (recipient of foodstamps, Medicaid, or WIC state-funded assistance), adequacy of prenatal care received, health-related behaviors (maternal tobacco or alcohol use), and presence of medical risk factors. Individual-level risk factors were selected from the data set based on clinical relevance and association with preterm birth9.

Maternal and paternal education levels are recorded in the database in years of education completed, which we dichotomized as educational levels of < 12 years versus 12 years or more in order to identify those with less than a high school education. The variable of maternal education had minimal missing data, and paternal education was less complete with 22.6% missing data. A composite dichotomous variable of low socioeconomic status was created from the individual dichotomous variables of recipient of any of three state funded support programs (Medicaid, foodstamps, and Special Supplemental Nutrition Program for Women, Infants and Children [WIC]). A dichotomous variable of inadequate prenatal care was created from a continuous variable which indicated the month of pregnancy when prenatal care was initiated. Inadequate prenatal care was defined as having initiated prenatal care after 20 weeks of pregnancy, which is the latter half of pregnancy. A composite variable of heterogeneous medical risk factors indicated pregnancies complicated by anemia (hematocrit < 30% or hemoglobin < 10 gm/dL), maternal cardiac disease, acute or chronic lung disease, diabetes (insulin dependent), diabetes (other), genital herpes, hydramnios/oligohydramnios, hemoglobinopathy, chronic hypertension, pre-eclampsia, eclampsia, incompetent cervix, previous infant weighing > 4000 gm, previous preterm or small-for-gestational-age infant, maternal renal disease, Rh sensitization, or uterine bleeding.

Poverty rate was obtained from US census data (1990) and was defined as the percentage of the population falling below the US federal poverty line at the county level of the mother's reported residence as a measure of area socioeconomic position. The poverty rate is a measure that is robust across various diseases and levels of geography; it has a link to possible policy implications, and is comparable over time1314. County-level poverty rates were divided into quartiles representing low poverty (first quartile, 0 – 6.86%), second quartile (6.89% – 11.91%), third quartile (11.92% – 14.49%) and high poverty (fourth quartile, ≥ 14.50%) to allow for nonlinear effects. The first quartile served as the reference group for comparison.

We used multilevel logistic regression analysis to estimate the effect of county-level poverty on preterm birth risk. This analytic method can estimate not only fixed effects of individual-level and area-level covariates, but also the random effect of geographic variation of preterm birth across counties. When significant geographic variation exists, this indicates that preterm birth was not randomly distributed. The 634,994 births were nested within 115 counties in Missouri (the City of St. Louis acts politically as a county).

Data were analyzed with SAS GLIMMIX macro (version 9.1, SAS Institute Inc., Cary, NC). Findings with a p-value of < 0.05 were considered statistically significant.

We calculated median odds ratios (MOR) and interval odds ratios (IOR)1516 to estimate the effect of cluster-level (level 2) factors on preterm birth via several successive regression models (Models I through V). The methods for calculating MOR and IOR have been previously described and are directly comparable with the fixed-effects odds ratios1516. MORs were calculated using the following equation:

M O R = exp ⁡ ( Z 0.75 * 2 * V C ) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemyta0Kaem4ta8KaemOuaiLaeyypa0JagiyzauMaeiiEaGNaeiiCaa3aaeWaaeaacqWGAbGwdaWgaaWcbaGaeGimaaJaeiOla4IaeG4naCJaeGynaudabeaakiabcQcaQmaakaaabaGaeGOmaiJaeiOkaOIaemOvay1aaSbaaSqaaiabdoeadbqabaaabeaaaOGaayjkaiaawMcaaaaa@40CA@

where Z_0.75 is the 75th percentile of standard normal distribution, V_C is the variance of PTB between counties, and exp(·) is the exponential function15. If the MOR is equal to 1, there is no variation between counties (no level 2 variation), but it is large if considerable intra-county variation exists.

80%-IORs were computed using the following equation:

I O R L o w e r = exp ⁡ ( β + Z 0.10 * 2 * V C ) , and I O R U p p e r = exp ⁡ ( β + Z 0.90 * 2 * V C ) MathType@MTEF@5@5@+=feaagaart1ev2aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@6D2F@

where Z_0.10 and Z_0.90 are the values of the standard normal distribution at the 10th and 90th percentiles, respectively. β is the regression coefficient of each category of county-level poverty rate. Larger geographic variation results in a broader range of the IOR. If the IOR does not cross the value of one, it suggests that county-level poverty substantially contributed to the geographic heterogeneity of PTB.

The study population was comprised of 634,994 live births to mothers who resided in 115 counties in Missouri from 1989 – 1998 for an average of 5,522 births per county (range: 156 – 119,244). The racial distribution of the study cohort was 82.4% White, 16.1% Black, 0.3% Indian, 0.5% Asian/Pacific Islander, 0.6% other/unknown.

Table 1 demonstrates the mothers' baseline demographic, medical and obstetric characteristics according the quartile of poverty rate of the county in which they resided at the time of the birth. Women who resided in counties with higher rates of poverty were significantly younger, and more likely to be black, less likely to graduate from high school, to be unmarried, and be of low income. Likewise, parturiants residing in areas with higher county-level poverty were more likely to have inadequate prenatal care, smoke cigarettes, drink alcohol during pregnancy, and have at least one medical risk factor.

The overall rate of preterm birth < 35 weeks for the entire study population was 3.9%. The rate of preterm birth < 35 weeks was 2.8% for mothers who resided in counties within the lowest quartile of poverty (Table 2). The rate of preterm birth increased with increasing county poverty rate (2^nd quartile: 3.4%, 3^rd quartile: 3,9%, 4^th quartile: 4.9%), p < 0.0001. Similarly, early preterm births (< 32 weeks) occurred more commonly in mothers who resided in counties with higher poverty rates. The rate of early preterm birth nearly doubled from 1.0% in the 1^st quartile to 1.9% in the quartile with the highest poverty rate (p < 0.0001).

Compared to births to mothers residing in counties with the lowest poverty rate (1^st quartile), the risk of preterm birth < 35 weeks increased with increasing rates of county-level poverty in univariate analysis (Table 3, Model I). This risk increase resulted in women in counties with the highest poverty rate being 1.30 times more likely to deliver preterm. The risk of early preterm birth was also significantly increased in mothers living in areas with the highest county-level poverty rate (OR 1.40; 95% CI 1.14, 1.72). This effect was similar when the study population was stratified by race. Both black and white mothers residing in counties with the highest poverty rate had an increased risk of preterm birth (Table 4).

When demographic factors are added to county-level poverty (Table 3, Model II), the odds ratio for preterm birth < 35 weeks of women residing in counties with the highest poverty rate was reduced from 1.30 to 1.15. Subsequent addition of other individual-level factors to the model, as demonstrated in models III, IV, and V, had minimal influence on the risk of preterm birth < 35 weeks (ORs ranging from 1.15–1.18) compared to the risk when only demographic factors are considered (4^th quartile, OR 1.15, 95% CI 1.09, 1.56).

When considering preterm births that occurred at earlier gestational ages (< 32 weeks of gestation), the risk increase associated with high county-level poverty was similarly influenced by individual-level factors. The unadjusted risk of early preterm birth with high county-level poverty (4^th quartile, OR: 1.40) was reduced to 1.0 when individual-level demographic factors were added to the model. The addition of other individual-level covariates such as prenatal care, behavioral risk factors, and medical risk factors only minimally influence the adjusted odds ratio (range 1.23 to 1.27).

Stratified analysis by race (black, white) demonstrated similar effect sizes when individual-level factors were added to the models (Table 4). Many more of the confidence intervals of risk in the analysis stratified by black race cross one than in the analysis by white race, but the effect sizes are very similar regardless of race.

There was a modest amount of geographic heterogeneity of PTB across counties in Missouri as measured by MOR and IOR. In all multilevel models, MOR values ranged from 1.07 to 1.22, and IOR-80s ranged from 0.79 to 2.05. As some IOR ranges did not contain one, this suggests that county-level poverty was an important area-level characteristic for geographic heterogeneity of PTB. But, other unmeasured area-level chacteristics could have also contributed to this geographic outcome variation.