Infant Mortality in Rural and Nonrural Counties in the United States

We conducted a cohort study of all live US births from 2014 to 2016 using 3 sources of data, linked to individual records using federal information processing codes for the mother’s resident county. We used 2014 to 2017 geocoded Period Linked Birth/Infant Death Data Sets (henceforth “birth and death records”) containing county of residence, 2010 to 2014 American Community Survey county-level 5-year samples, and 2010 to 2014 data included in the 2016 to 2017 Area Health Resource File publicly available from the Health Services Resources and Services Administration.²⁰ 

The birth and death records were provided with permission from the National Association for Public Health Statistics and Information Systems as compiled from data provided by the 57 vital statistics jurisdictions through the Vital Statistics Cooperative Program. These records link birth and death certificates of each infant dying within the first year after birth. The recent rate of record linkage between birth records and infant death records is excellent. In the 2014 linked file, only 161 of 23 247 infant deaths were unlinked; similarly, in the 2015 linked file, 143 of 23 490 infant deaths were unlinked, and in 2016, only 113 of 23 079 were unlinked.^21–23  Each birth record includes the federal information processing code of the mother’s county of residence, which we use to link the birth and death records to the socioeconomic, demographic, and health care access data from the American Community Survey and Area Health Resource File.

We constructed cohort files from the 2014 to 2017 period files for use in this analysis, the most current available at the start of this study, because the last cohort file produced by the National Center for Health Statistics (NCHS) was for 2013 births. We used the DUPLICATE command in Stata to identify infant deaths that occurred in the period file of the subsequent year. Using this process, we identified and replaced all but 3 birth records with the linked birth–infant death files for 2014, 2015, and 2016, thus creating cohort files from the period files.

From a total population of 11 912 448 US resident births, we excluded all births occurring before 20 weeks, births occurring after 44 weeks, and births in which the gestational age–specific birth weight is >4 SD from the mean (n = 30 137 cases). Records missing information on resident county (n = 5 counties and n = 497 live births) were also excluded, leaving 11 881 814 observations, 99.7% of the national cohort of US resident births.

The records for 2014 to 2016 reflect a transition to the 2003 Revised Certificate of Live Birth and thus include data from both the revised and the unrevised 1989 versions.²⁴  For the descriptive analysis and the model that included individual-level data, we excluded Connecticut, Rhode Island, and New Jersey (n = 449 739 births) because they had not yet adopted the 2003 Revised Certificate of Live Birth; also Georgia, Virginia, and Hawaii (n = 756 484), which had a substantial fraction of missing data for key covariates, were excluded in some analyses. Cases with missing data (n = 622 939 births) were also excluded from the analyses. After exclusions, 10 109 644 births, or 84.9% of total US live births, were included in the analyses that included individual-level characteristics.

The primary outcome was infant death, within the first year of life overall and during the neonatal (birth to <28 days) and postneonatal periods (28 days to 1 year), the latter assessed among those infants who had survived the neonatal period, as identified in the linked birth and death records. The primary exposure was a county-level measure of rurality and urbanicity of the mother’s resident county by using the NCHS urban-rural classification scheme, which categorizes areas by metropolitan or micropolitan classification, population size, and whether the county is part of the metropolitan core.²⁵  Figure 1 shows a map of the 6 NCHS categories: large central metropolitan, large fringe metropolitan, median metropolitan, small metropolitan, micropolitan, and noncore (rural). In some classification analyses, micropolitan and noncore counties are combined into a single rural category. In these analyses, we chose to consider them as separate groups. To account for regional effects and spatial autocorrelation, we used the 10 US Department of Health and Human Services Regions and also the longitude and latitude of the county geometric centroid from the ArcGIS shapefiles.

We used 2 county-level measures reflecting access to perinatal care: dichotomous variables indicating counties with a hospital that offers obstetrical care and counties with a hospital with a NICU. In addition, we included more general factors related to health care by using dichotomous variables indicating whether a county has been designated as a Health Professional Shortage Area (HPSA) for primary care or had a Federally Qualified Health Center (FQHC) or Rural Health Clinic (RHC) in the county. These measures are included to reflect access to health care before pregnancy and infant and pediatric care.

We estimated county-level socioeconomic status using the “advantage index” developed by Diez Roux et al.²⁶  This index includes 5 components: logged median household income; logged median value of owner-occupied housing units; the percentage of households with interest, dividend, or net rental income; the percentage of workers in management, science, and arts occupations; and the percentage of residents >25 with at least a high school diploma. We calculated this index by summing the z scores for each component as previously described.²⁶  Positive values on the scale indicate socioeconomic advantage, whereas negative values indicate disadvantage.

We also included several county-level measures of population characteristics: fraction of county-level population by race using non-Hispanic (NH) white, NH Black, Hispanic, or NH other. Furthermore, because the birth cohort reflects the characteristics of the population, including local norms of family size, which are known to differ by geography, we included county-level total fertility rate (TFR).²⁷ 

Last, we examined available individual-level factors known to be associated with infant mortality, including mother’s demographic characteristics, tobacco use, diagnosis of gestational diabetes, diabetes mellitus, chronic hypertension, gestational hypertension and preeclampsia, prepregnancy BMI, method of delivery and whether at a hospital, and whether delivery was performed by a physician. Finally, we also included infant sex and plurality.

We calculated infant mortality rates across the 6 NCHS categories for the 3 years of pooled births as deaths per 1000 live births, considering overall infant mortality as well as neonatal and postneonatal mortality. We conducted a descriptive analysis comparing the county and individual-level characteristics of women across the 6 NCHS categories. Analysis of variance and χ² tests were used to test the significance of differences across geographic classification, with P < .05 considered to be significant. We also examined associations of each component of the advantage index with the outcome and in the full structural models to determine if one or more components were driving the observed associations.

We used generalized linear models to specify multilevel analyses using county fixed effects while allowing individual-level factors to vary. Because of the substantial regional variation in the distribution of rural and urban counties across the United States, we control both for region of the United States and for location using latitude–longitude coordinates for county centroids. Because the Index of Dissimilarity was not independently associated with the outcome after controlling for county-level population characteristics, it was not included in the models.

• Model 1 = NCHS category plus county centroids and regional controls.

• Model 2 = model 1 plus county-level health care factors (HPSA, obstetric hospital, NICU, FQHC/RHC).

• Model 3 = model 1 plus county-level socioeconomic status (advantage index).

• Model 4 (full structural model) = model 1 plus county-level socioeconomic status (advantage index) plus county population characteristics (fraction white, Black, Hispanic, other; TFR).

• Model 5 (complete model) = model 4 plus individual-level variables.

We used the Bayesian information criterion (BIC) measure to compare model fit. Sensitivity analyses was conducted to ensure excluding cases with missing data in model 5 did not change our findings. Sensitivity analyses included modeling of 2016 alone, the year in which all states had adopted the new birth certificate. Because we were unable to identify siblings of multiple gestation pregnancies in the data, we also ran separate models that included only singleton births to account for the potential effects of nonindependence.

All analyses were conducted by using Stata SE V16.0 (Stata Corp, College Station, TX).²⁸ 

The sample included 11 881 814 live births in 2014 to 2016 and 64 478 infant deaths distributed across 3134 counties (Table 1). The infant mortality rate was lowest among infants born to mothers living in large fringe metropolitan counties. There were significant differences in the odds of infant death when comparing with fringe metropolitan counties, controlling for year, county location, and region. As counties become more rural, the odds increased and were greatest for infants of women living in noncore counties, with an odds ratio (OR) of 1.32 (95% confidence interval [CI]: 1.26–1.39) when compared with large fringe metropolitan counties. Rural-urban mortality gradients varied by period of infant death, with odds of neonatal death lowest in large fringe metropolitan counties and an inconsistent trend with increasing rurality. This contrasts with postneonatal mortality, in which the odds were progressively higher as counties became more rural, with an OR of 1.57 (95% CI: 1.46–1.69) for noncore counties when compared with large fringe metropolitan counties.

County-level characteristics (Table 2) varied by rural-urban classification (all P < .001) in measures that reflect access to health care resources, including the presence of an obstetric hospital, a facility with a NICU, a FQHC/RHC, or whether the county is identified as a HPSA. For example, whereas 100% of large central metropolitan counties had ≥1 obstetric hospital in that county, 68.2% of micropolitan and only 25.7% noncore counties had any obstetric hospital. Differences were even greater in the percentages of counties with ≥1 hospital with an NICU. Although 97.1% of large central metropolitan counties had an NICU, only 6.1% of micropolitan and 1.4% noncore counties had an NICU. There were also differences in socioeconomic status as measured by using the advantage index, as well as characteristics of the population. All individual-level characteristics (Table 3), with the exception of infant sex, also differed across the rural-urban continuum including age at delivery, completed education, BMI, and tobacco use (P < .001).

The results of the multilevel models for overall mortality, neonatal mortality, and postneonatal mortality are presented in Table 4. Regional controls did little to change to the varying odds across rural-urban classification (model 1). The presence of an obstetric hospital and an NICU was significantly associated with higher odds of infant mortality overall and for both neonatal and postneonatal mortality (model 2), whereas county designation as HPSA or with an FQHC/RHC was not. Including health system characteristics in the model did little to attenuate the differences in odds of infant mortality across rural and urban classification groups.

In contrast, the advantage index was positively associated with infant mortality (model 3), overall as well as neonatal and postneonatal, and when included in the model eliminated any significant difference in the odds of infant mortality between the micropolitan and noncore counties and the large metropolitan fringe classification groups for both overall and neonatal mortality outcomes. For postneonatal mortality, including the advantage index eliminated the differences between micropolitan counties, and only a small difference remained for noncore counties. In addition, this also eliminated any association between the health care characteristics of the county with infant mortality. Analyses of the individual components of the advantage index reveal these associations were largely driven by household income and median value of owner-occupied housing units, each inversely related to the odds of an infant death (Supplemental Table 5).

County population-level characteristics were included in the full structural model (Table 4, model 4). County race and ethnicity were associated with the outcome (P < .0001 for each) but had little effect on the overall differences across the rural-urban classification groups. The TFR was not independently associated with the odds of infant mortality. When structural factors and individual-level characteristics were included in the full model (Table 4, model 5), differences by county-level rural-urban geographic classification were small and significant only for medium and small metropolitan counties. There were no differences by region, and the population level race and ethnic composition and advantage index were each significantly associated with the outcome (all P < .001). Including individual-level characteristics, such as type of medical insurance and health behaviors such as tobacco use, did little to change the county-level odds. Results from each model, including all individual-level covariates, are included in the Supplemental Tables 6 through 8.

Separate sensitivity analyses using data for 2016 only, and for singleton births, revealed no meaningful differences in the associations found for the entire cohort.

Using the BIC to assess model fit, we found the addition of health care factors (model 2) led to a poorer fit, whereas fit improved for the other models. The advantage index improved the postneonatal model more than the neonatal model. The model predicting postneonatal mortality had a substantially better fit, based on the BIC, than those predicting overall and neonatal mortality.

In this national study of all 2014 to 2016 US resident births, rates of infant mortality were higher as counties became more rural and driven by higher rates of mortality in the postneonatal period. Although individuals living in rural counties were less likely to have access to an obstetric hospital or an NICU in their county, with our multilevel analysis we found no evidence that this factor explains the greater odds of infant mortality in rural counties. The more prevalent use of tobacco, presence of other health risks, and lower rates of private insurance also did not account for the differences. Instead, we find the higher infant mortality rates in rural counties are best explained by the greater socioeconomic disadvantage of their county of residence.

The association of socioeconomic factors with mortality is consistent with United States and international studies of mortality.^5,15,29–32  Although access to health care may be of unique relevance to maternal and infant health, our analysis finds little independent contribution of local health care resources to the higher rates of infant mortality in rural counties in the United States, net of socioeconomic and individual factors. The resources of hospitals with higher levels of care, including NICUs, are critically important to improved survival of infants born at early gestational ages. However, our results are consistent with a growing literature showing infant mortality rates due to congenital malformations, sudden unexpected infant death as well as higher rates of unintentional injuries are higher in more rural counties, and not low birth weight or maternal complications.^2,33 

As in previous literature, we found that individuals living in rural counties are less likely to live in a county with an obstetric hospital or an NICU. However, these resources were not independently associated with odds of infant mortality. This may be due to the success of formal and informal systems of regionalized perinatal care supporting maternal and infant health, connecting more rural communities with larger urban centers offering higher levels of obstetrical and neonatal intensive care services.^34,35  Alternatively, it may reflect the more limited impact of these systems on this important outcome.³²  This is especially likely when considering their limited impact on postneonatal deaths because their cause of death is unlikely to be related to NICU care. Also consistent with previous research, many individual-level health risk factors, including lack of private insurance, tobacco use, and obesity, were more common as counties became more rural. However, their greater prevalence did not explain the higher infant mortality rates in micropolitan and noncore counties.

Limitations of this study are important to consider. The NCHS rural-urban classification of counties is a crude assessment of rurality and does not fully account for within-county variation or resources in nearby metropolitan areas. Similarly, the county-level data used to estimate socioeconomic resources may not reflect the spatial variability across neighborhoods and communities, but small-area data are difficult to examine on the national scale and for rare outcomes such as infant death. As in any study of this magnitude, missing data are inevitable. Some states were dropped from our final model; however, sensitivity analyses suggest their exclusion made little difference in the results. In addition, the results do not address the potential association of these factors with stillbirth or perinatal mortality, an important area for future research.

Differences in infant mortality may be better understood within a multilevel framework informed by the socio-ecological perspective and considering the importance of socioeconomic factors. Despite increasing levels of urbanization, 13.5% of 2014 to 2016 live births in the US occurred to residents of micropolitan and noncore (rural) counties. Given national goals to reduce infant mortality and achieve health equity in birth outcomes, health policymakers should also pay attention to the differing outcomes and socioeconomic milieux along the rural-urban continuum. These findings suggest that investment beyond medical care may be needed if we are to continue to see improvements in infant mortality at the national level and equity in rural areas.

We thank Dr Emily Shrider for her valuable contributions to early versions of this article. This research uses data from the NCHS with permission from the National Association for Public Health Statistics and Information Systems as compiled from data provided by the 57 vital statistics jurisdictions through the Vital Statistics Cooperative Program. All other data are from publicly available sources.

BIC

Bayesian information criterion

CI

confidence interval

FQHC

Federally Qualified Health Center

HPSA

Health Professional Shortage Area

NCHS

National Center for Health Statistics

NH

non-Hispanic

OR

odds ratio

RHC

Rural Health Clinic

TFR

total fertility rate