Bias correction models for electronic health records data in the presence of non-random sampling.

Electronic health records (EHRs) contain rich clinical information for millions of patients and are increasingly used for public health research. However, non-random inclusion of subjects in EHRs can result in selection bias, with factors such as demographics, socioeconomic status, healthcare referral patterns, and underlying health status playing a role. While this issue has been well documented, little work has been done to develop or apply bias-correction methods, often due to the fact that most of these factors are unavailable in EHRs. To address this gap, we propose a series of Heckman type bias correction methods by incorporating social determinants of health selection covariates to model the EHR non-random sampling probability. Through simulations under various settings, we demonstrate the effectiveness of our proposed method in correcting biases in both the association coefficient and the outcome mean. Our method augments the utility of EHRs for public health inferences, as we show by estimating the prevalence of cardiovascular disease and its correlation with risk factors in the New York City network of EHRs.

Effect of health systems context on infant and child mortality in sub-Saharan Africa from 1995 to 2015, a longitudinal cohort analysis.

Each year, > 3 million children die in sub-Saharan Africa before their fifth birthday. Most deaths are preventable or avoidable through interventions delivered in the primary healthcare system. However, evidence regarding the impact of health system characteristics on child survival is sparse. We assembled a retrospective cohort of > 250,000 children in seven countries in sub-Saharan Africa. We described their health service context at the subnational level using standardized surveys and employed parametric survival models to estimate the effect of three major domains of health services-quality, access, and cost-on infant and child survival, after adjusting for child, maternal, and household characteristics. Between 1995 and 2015 we observed 13,629 deaths in infants and 5149 in children. In fully-adjusted models, the largest effect sizes were related to fees for services. Immunization fees were correlated with poor child survival (HR = 1.20, 95% CI 1.12-1.28) while delivery fees were correlated with poor infant survival (HR = 1.11, 95% CI 1.01-1.21). Accessibility of facilities and greater concentrations of private facilities were associated with improved infant and child survival. The proportion of facilities with a doctor was correlated with increased risk of death in children and infants. We quantify the impact of health service environment on survival up to five years of age. Reducing health care costs and improving the accessibility of health facilities should remain a priority for improving infant and child survival. In the absence of these fundamental investments, more specialized interventions may not achieve their desired impact.

Residential Racial Isolation and Spatial Patterning of Type 2 Diabetes Mellitus in Durham, North Carolina.

Neighborhood characteristics such as racial segregation may be associated with type 2 diabetes mellitus, but studies have not examined these relationships using spatial models appropriate for geographically patterned health outcomes. We constructed a local, spatial index of racial isolation (RI) for black residents in a defined area, measuring the extent to which they are exposed only to one another, to estimate associations of diabetes with RI and examine how RI relates to spatial patterning in diabetes. We obtained electronic health records from 2007-2011 from the Duke Medicine Enterprise Data Warehouse. Patient data were linked to RI based on census block of residence. We used aspatial and spatial Bayesian models to assess spatial variation in diabetes and relationships with RI. Compared with spatial models with patient age and sex, residual geographic heterogeneity in diabetes in spatial models that also included RI was 29% and 24% lower for non-Hispanic white and black residents, respectively. A 0.20-unit increase in RI was associated with an increased risk of diabetes for white (risk ratio = 1.24, 95% credible interval: 1.17, 1.31) and black (risk ratio = 1.07, 95% credible interval: 1.05, 1.10) residents. Improved understanding of neighborhood characteristics associated with diabetes can inform development of policy interventions.

A retrospective cohort study to quantify the contribution of health systems to child survival in Kenya: 1996-2014.

Globally, the majority of childhood deaths in the post-neonatal period are caused by infections that can be effectively treated or prevented with inexpensive interventions delivered through even very basic health facilities. To understand the role of inadequate health systems on childhood mortality in Kenya, we assemble a large, retrospective cohort of children (born 1996-2013) and describe the health systems context of each child using health facility survey data representative of the province at the time of a child's birth. We examine the relationship between survival beyond 59 months of age and geographic distribution of health facilities, quality of services, and cost of services. We find significant geographic heterogeneity in survival that can be partially explained by differences in distribution of health facilities and user fees. Higher per capita density of health facilities resulted in a 25% reduction in the risk of death (HRR = 0.73, 95% CI:0.58 to 0.91) and accounted for 30% of the between-province heterogeneity in survival. User fees for sick-child visits increased risk by 30% (HRR = 1.30, 95% CI:1.11 to 1.53). These results implicate health systems constraints in child mortality, quantify the contribution of specific domains of health services, and suggest priority areas for improvement to accelerate reductions in child mortality.

Assessing Geographic Variation in Strabismus Diagnosis among Children Enrolled in Medicaid.

PURPOSE: To determine how strabismus diagnosis varies within a given community and across communities among children with Medicaid health insurance. DESIGN: Retrospective cohort analysis. PARTICIPANTS: Children aged ≤10 years enrolled in Medicaid in Michigan or North Carolina during 2009. METHODS: Children who met the study inclusion criteria were identified from the Medicaid Analytic Extract database, which includes claims data for all children enrolled in Medicaid throughout the United States. Residential location was determined by the last known 5-digit ZIP code for each child, which was linked to the centroid of a ZIP Code Tabulation Area (ZCTA) for geo-referencing and spatial analyses. International Classification of Diseases, 9th Revision, Clinical Modification billing codes were used to identify children diagnosed with strabismus (code 378.xx). Bayesian hierarchical intrinsic conditional autoregressive spatial probit models were used to determine the risk of a child receiving a strabismus diagnosis in communities throughout Michigan and North Carolina. Maps display communities (ZCTAs) where the 95% credible intervals for the spatial random effects estimates do not cross zero, allowing for identification of locations with increased and decreased strabismus diagnosis risk relative to other communities in the states. MAIN OUTCOME MEASURES: Likelihood of receiving a diagnosis of strabismus. RESULTS: In 2009, among 519 212 eligible children in Michigan, 7535 (1.5%) received ≥1 strabismus diagnosis, and in North Carolina, 5827 of 523 886 eligible children (1.1%) were diagnosed with strabismus. In both states, the proportion receiving a strabismus diagnosis among black (0.9% in Michigan; 0.7% in North Carolina) and Hispanic (1.1% in Michigan; 0.8% in North Carolina) children was lower than the proportion for white children (1.8% in Michigan; 1.6% in North Carolina). Children living in poorer communities in both states were less likely to be diagnosed with strabismus independent of their race/ethnicity. CONCLUSIONS: A child's likelihood of being diagnosed with strabismus is associated with characteristics of the residential community where he or she resides. The findings of this study highlight the importance of ensuring that children who live in less affluent communities have access to the necessary services and eye care professionals to properly diagnose and treat them for this condition.

Racial isolation and exposure to airborne particulate matter and ozone in understudied US populations: Environmental justice applications of downscaled numerical model output.

BACKGROUND: Researchers and policymakers are increasingly focused on combined exposures to social and environmental stressors, especially given how often these stressors tend to co-locate. Such exposures are equally relevant in urban and rural areas and may accrue disproportionately to particular communities or specific subpopulations. OBJECTIVES: To estimate relationships between racial isolation (RI), a measure of the extent to which minority racial/ethnic group members are exposed to only one another, and long-term particulate matter with an aerodynamic diameter of <2.5µ (PM2.5) and ozone (O3) levels in urban and nonurban areas of the eastern two-thirds of the US. METHODS: Long-term (5year average) census tract-level PM2.5 and O3 concentrations were calculated using output from a downscaler model (2002-2006). The downscaler uses a linear regression with additive and multiplicative bias coefficients to relate ambient monitoring data with gridded output from the Community Multi-scale Air Quality (CMAQ) model. A local, spatial measure of RI was calculated at the tract level, and tracts were classified by urbanicity, RI, and geographic region. We examined differences in estimated pollutant exposures by RI, urbanicity, and demographic subgroup (e.g., race/ethnicity, education, socioeconomic status, age), and used linear models to estimate associations between RI and air pollution levels in urban, suburban, and rural tracts. RESULTS: High RI tracts (≥80th percentile) had higher average PM2.5 levels in each category of urbanicity compared to low RI tracts (<20th percentile), with the exception of the rural West. Patterns in O3 levels by urbanicity and RI differed by region. Linear models indicated that PM2.5 concentrations were significantly and positively associated with RI. The largest association between PM2.5 and RI was observed in the rural Midwest, where a one quintile increase in RI was associated with a 0.90µg/m(3) (95% confidence interval: 0.83, 0.99µg/m(3)) increase in PM2.5 concentration. Associations between O3 and RI in the Northeast, Midwest and West were positive and highest in suburban and rural tracts, even after controlling for potential confounders such as percentage in poverty. CONCLUSION: RI is associated with higher 5year estimated PM2.5 concentrations in urban, suburban, and rural census tracts, adding to evidence that segregation is broadly associated with disparate air pollution exposures. Disproportionate burdens to adverse exposures such as air pollution may be a pathway to racial/ethnic disparities in health.

Racial residential segregation and preterm birth: built environment as a mediator.

BACKGROUND: Racial residential segregation has been associated with preterm birth. Few studies have examined mediating pathways, in part because, with binary outcomes, indirect effects estimated from multiplicative models generally lack causal interpretation. We develop a method to estimate additive-scale natural direct and indirect effects from logistic regression. We then evaluate whether segregation operates through poor-quality built environment to affect preterm birth. METHODS: To estimate natural direct and indirect effects, we derive risk differences from logistic regression coefficients. Birth records (2000-2008) for Durham, North Carolina, were linked to neighborhood-level measures of racial isolation and a composite construct of poor-quality built environment. We decomposed the total effect of racial isolation on preterm birth into direct and indirect effects. RESULTS: The adjusted total effect of an interquartile increase in racial isolation on preterm birth was an extra 27 preterm events per 1000 births (risk difference = 0.027 [95% confidence interval = 0.007 to 0.047]). With poor-quality built environment held at the level it would take under isolation at the 25th percentile, the direct effect of an interquartile increase in isolation was 0.022 (-0.001 to 0.042). Poor-quality built environment accounted for 35% (11% to 65%) of the total effect. CONCLUSION: Our methodology facilitates the estimation of additive-scale natural effects with binary outcomes. In this study, the total effect of racial segregation on preterm birth was partially mediated by poor-quality built environment.

A spatial bivariate probit model for correlated binary data with application to adverse birth outcomes.

Motivated by a study examining geographic variation in birth outcomes, we develop a spatial bivariate probit model for the joint analysis of preterm birth and low birth weight. The model uses a hierarchical structure to incorporate individual and areal-level information, as well as spatially dependent random effects for each spatial unit. Because rates of preterm birth and low birth weight are likely to be correlated within geographic regions, we model the spatial random effects via a bivariate conditionally autoregressive prior, which induces regional dependence between the outcomes and provides spatial smoothing and sharing of information across neighboring areas. Under this general framework, one can obtain region-specific joint, conditional, and marginal inferences of interest. We adopt a Bayesian modeling approach and develop a practical Markov chain Monte Carlo computational algorithm that relies primarily on easily sampled Gibbs steps. We illustrate the model using data from the 2007-2008 North Carolina Detailed Birth Record.