Study sample

Delivery hospitalizations were identified based on International Classification of Diseases–ninth revision-Clinical Modification diagnosis and procedure codes and diagnosis-related group delivery codes. We examined all deliveries that occurred in hospitals with at least 10 deliveries annually because we wanted to analyze hospitals with an obstetric volume that was high enough to preclude hospitals with no routine obstetrics practice and therefore excluded 408 (77 unweighted) deliveries at 171 hospitals (35 unweighted). This resulted in the exclusion of 0.005% white deliveries and 0.007% black deliveries. We limited our analyses to black and white deliveries and excluded 2,472,365 additional deliveries on this basis. Some hospitals do not record data on race/ethnicity routinely in their hospital discharge summaries; thus, race was missing for 722 hospitals (152 unweighted), which was approximately 11.2% of the hospitals. Severe maternal morbidity was not different for deliveries in the hospitals with and without data on race (1.51 vs 1.34; P = .37). Our analysis thus included 5539 hospitals (1146 unweighted, 1140 unique hospitals for combined 2010-2011 dataset) that provided delivery services and 4,609,291 deliveries (942,622 unweighted). Only national estimates are presented in the results section.

Severe maternal morbidity

We used a published algorithm defined by investigators from the Centers for Disease Control and Prevention to identify severe maternal morbidity (primary outcome), which is a potentially life-threatening diagnosis or receipt of a life-saving procedure (eg, renal failure, shock, embolism, eclampsia, septicemia, mechanical ventilation, transfusion). The algorithm includes 25 categories that capture indicators of organ-system dysfunction. It hierarchically classifies all severe maternal morbidity hospitalizations that are associated with in-hospital death or transfer because of hospitalizations with severe complications, regardless of the length of stay. In-hospital death was identified with the use of the variable “died during hospitalization” and transfer status using “disposition of patient” or “admission source” in the Nationwide Inpatient Sample. Length of stay is also a part of this algorithm. Hospitalizations with severe morbidity rates and a short length of stay were not classified as cases of severe maternal morbidity as specified by the Centers for Disease Control and Prevention algorithm. Short length of hospital stay was defined as length of stay <90th percentile as calculated separately for vaginal, primary, and repeat cesarean deliveries. Length of stay restrictions were not applied to delivery hospitalizations with severe complications that were identified by procedure codes (eg, hysterectomy, blood transfusion, ventilation), as recommended.

Black-serving delivery hospitals

Similar to methods used by Jha et al and Lopez and Jha, we ranked hospitals by their proportion of black deliveries among all deliveries and chose 2 cut-off points. We defined the top 5% of hospitals as high black-serving hospitals, the next 20% (those in the >5% to ≤25% range) as medium black-serving hospitals, and the remaining 75% of hospitals as low black-serving hospitals. Two hundred seventy-nine hospitals (5%) were designated as high black-serving; 1106 hospitals (20%) were designated as medium black-serving, and 4102 hospitals (75%) were designated as low black-serving.

Covariables

Risk adjustment variables were chosen by their association with the outcome, biologic plausibility, and previous research. Our aim was to adjust for factors that were likely to affect maternal morbidity and hospital of delivery. We included maternal sociodemographic characteristics (age, race, and income [the latter broken down by zip code] ), clinical factors (eg, multiple births, previous cesarean delivery) and comorbid conditions and antepartum complications (eg, diabetes mellitus, hypertension, preeclampsia, obesity, preterm labor, premature rupture of membranes, disorders of placentation). Hospital characteristics included teaching status, hospital ownership, delivery volume, rural or urban location, hospital region (Northeast, Midwest, South, and West), bed size, and percentage of Medicaid admissions as a proxy for the proportion of poor patients that the hospitals serve.

Analysis

We compared the characteristics of black vs white women with severe maternal morbidity with Rao-Scot chi square tests for categoric data. We used analysis of variance and chi square tests, as appropriate, to examine the characteristics of delivery hospitals according to the proportion of their black deliveries. For our primary outcome, risk-adjusted probability of severe maternal morbidity, we estimated 4 multivariable patient-level logistic regression models. Each model incorporated the survey weights; the variance estimates accounted for the Nationwide Inpatient Sample structure.

The first model included race and patient risk factors only; the second included these variables plus indicator variables for whether the hospital was a high, medium, or low black-serving hospital. The third included these variables plus the other hospital levels variables listed earlier. For the fourth model, we included patient risk factors, race, and black-serving status of the hospital and variables that represented the interaction between race and black-serving status of the hospital. Using this model, we estimated the risk-adjusted predicted probability of severe maternal morbidity for 6 categories of race and hospital: black and white patients, respectively, at high black-serving hospitals; black and white patients, respectively, at medium black-serving hospitals; and black and white patients, respectively, at low black-serving hospitals. We used recycled predictions to calculate predicted probabilities using the margins command; confidence intervals were estimated with the use of the delta method.

We conducted 2 sensitivity analyses. First, we estimated the multivariable patient-level logistic regression model and added cesarean delivery as a covariate. Second, we conducted the analyses after excluding hospitals with <30 deliveries. All statistical analyses were performed with SAS system software (version 9.3; SAS Institute Inc, Cary, NC) and Stata/SE (version 13.1; StataCorp LP, College Station, TX). All tests were 2-tailed; a probability value of .05 was considered statistically significant.