Data

Data were pooled from multiple sources, including birth certificate data from the Centers for Disease Control and Prevention (CDC), hospital data from the American Hospital Association (AHA) survey, and spending and usage data from the Dartmouth Atlas. The Dartmouth Atlas uses Hospital Service Areas (HSAs) to define local markets for hospital-based medical care and as a way to measure the variation in primarily hospital-based services (such as cesarean deliveries) across US communities. All analyses were done at the HSA-level (ie, with each HSA as the unit of observation). We used individual birth certificate data from the CDC with the county of birth identified. These individual level data were aggregated to the HSA-level; county-level data were converted to zip code–level (weighted by percent of population in the zip code in each county, which was available from the US postal service), then zip code-level data were aggregated to the HSA-level (weighted by the number of each births that occurred in each zip code within that HSA). Data on area demographics were obtained from the Health Resources and Services Administration 2010 Area Resource File, which provide data at the county-level on basic demographics, such as physician density and median income, which was converted similarly to HSA-level data, as mentioned earlier. Hospital data were obtained from the 2010 AHA survey. The AHA survey provides basic information about hospitals, such as bed number and teaching status, which was aggregated to the HSA-level (the hospital HSA was available in the AHA survey). Area obesity rates were obtained from the 2010 National Health and Nutrition Examination Survey at the state-level and converted to the HSA-level. Similarly, state-level malpractice data from the 2010 National Practitioner Database were used. Medicare total spending amount and hospital days in the last 6 months of life were obtained at the HSA-level from the Dartmouth Atlas of Health Care, which derives this information from Medicare claims data.

Variables of Interest

Our primary outcome was variation in HSA-level total cesarean delivery rates. Because nulliparous, term, singleton, vertex (NTSV) cesarean delivery rates are a commonly defined subset to analyze cesarean delivery variation and are supported as a quality metric by the American College of Obstetricians and Gynecologists, we also looked at variation in this cohort. Our primary predictors were total spending per Medicare beneficiary and number of hospital days in the last 6 months of life per Medicare decedent that were chosen a priori as ways to measure broader health care spending and usage patterns of the HSA. We used 2010 Medicare total spending data (age, sex, and price-adjusted) and the last available year of end-of-life care data (2007).

In terms of covariates, we included birth certificate variables that were coded reliably and sufficiently prevalent in the population (at least 5%). Pregestational and gestational disease states for diabetes mellitus and hypertension were combined into larger categories of any diabetes mellitus or hypertension. Minor and major teaching hospital admissions were combined into a larger category of “any teaching” hospital. All malpractice claims across all specialties were used.

Analysis

First, we examined the overall variation in both total and NTSV cesarean delivery rates across HSAs. Then, we divided HSAs into those with low cesarean delivery rates (less than the median) and high cesarean delivery rates (more than the median) and examined differences in covariates between these HSAs. We calculated the correlation between HSA-level cesarean delivery rates and total Medicare spending per beneficiary as well as the correlation between hospital days in the last 6 months of life. For both of these correlations, we used Pearson coefficients, weighted by the number of births per HSA. We also performed a sensitivity analysis for these correlations by excluding outliers (HSAs <5th percentile or >95th percentile). Next, we divided HSAs into quartiles of Medicare spending and end-of-life care. We examined the unadjusted and adjusted rates of total and NTSV cesarean deliveries across these quartiles. For the adjusted rates, we used the least-squares means methods. All missing variables were input. We adjusted for maternal and birth characteristics (age, race, parity, gestational age, marital status, education, birthweight, multiple gestations, smoking, diabetes mellitus, hypertension, and breech presentation), area demographics (obesity, region in the United States, birth rate, and median income), hospital-level factors (teaching status, bed size, profit status), and malpractice (total claims and claims >$1 million). Obstetric practice variables (induction rate, forceps/vacuum rate, deliveries in the hospital, and provider type and density) were not adjusted for because they were believed potentially to be on the “causal pathway” for cesarean delivery. A sensitivity analysis was performed when the obstetric practice variables were included in the adjustment. Unadjusted and adjusted rates were compared across quartiles with analysis of variance testing. All analyses were weighted by the total number of deliveries per HSA.

To analyze the degree to which broader health care spending and usage measures explained the variation in cesarean delivery rates, we created separate linear regression models. Each model had HSA-level total or NTSV cesarean delivery as the outcome and different predictors for inputs. The R -squared value of that model was used to quantify the amount of variation in the HSA-level cesarean delivery rates that was explained by that variable or variable group. We adjusted for baseline differences in maternal and birth characteristics and area demographics using a partial R -squared model; the partial R -squared value represents the percent of the remaining variation (after accounting for baseline differences) that was explained by that variable or variable group.

For all analyses, SAS statistical software (version 9.3; SAS Institute Inc, Cary, NC) was used, and a probability value of <.05 was considered statistically significant. Our data were all deidentified and exempt from Institutional Review Board approval.