Approximately one in five births per year in the United States is a cesarean delivery, and nearly 60% of those are primary cesarean deliveries.1 There are considerable variations in cesarean delivery rates by geographic region, hospital, and type of practice.2–4 Although lowering the cesarean delivery rate might improve quality of health care and reduce expense, setting a target cesarean rate remains problematic. Ideally, target rate would help hospitals assess their performance and would be determined by studying which rates maximize maternal and neonatal outcomes. However, research on outcomes is hampered by complex interactions of maternal and newborn health issues. Without such data, different organizations have selected target rates by consensus; for example, 15% by the Healthy People 2000 initiative.5 Besides being arbitrary, the selection of a single low rate as a marker of quality for all hospitals does not account for hospital variations in maternal and neonatal risk.
An alternative approach to applying a single, arbitrary target rate for cesarean delivery to all hospitals is to compare hospital rates after adjusting for patient mix, taking into account the prevalence of risk factors that are strongly associated with cesarean deliveries.6 That approach uses average practice as the standard for comparison. At the risk of substituting a benchmark for true quality, ie, of accepting an average cesarean birth rate as appropriate, that approach might better identify hospitals to target for interventions.
The objective of this study was to determine whether risk adjusting cesarean delivery rates based on a hospital's case mix affected assessment of hospital performance. To accomplish that objective, we developed a method of adjusting primary cesarean delivery rates for patient mix by using birth certificate data from the state of Illinois. We used that model to make inferences about variations in rates among Illinois hospitals. Our analysis was limited to primary cesarean deliveries. Although the indications for cesarean delivery in a vaginal trial of labor after a prior cesarean delivery might be similar to those for a primary cesarean delivery, most repeat cesarean deliveries are done by patient choice without a trial of labor. Decisions for primary and repeat cesarean delivery are made differently and primary cesarean deliveries drive the total cesarean delivery rate. Thus, we chose to limit our study to primary cesarean deliveries.
The study population was derived from all deliveries in Illinois in 1991. Approximately 70% of deliveries occurred in the Chicago metropolitan area, a mix of inner-city and suburban communities, and the remainder occurred in downstate Illinois, which is predominately rural. The racial mix of women who delivered in the Chicago area was 75% white, 20% black, and 5% other, whereas downstate it was 90% white. All hospitals belong to one of ten perinatal networks, and about 1% of the obstetric population is transferred before delivery. The statewide primary cesarean delivery rate for 1991 was 15.1%, similar to the 1991 national rate of 15.9%.1
Birth certificate data in Illinois were collected from questionnaires and information generated from parental interviews, along with reviews of hospital charts by trained personnel. The birth certificate database contained parental demographics, maternal medical complications, month prenatal care began, type of delivery, hospital (coded for anonymity), infant gestational age, birth weight, gender, birth order, and congenital anomalies.
Data were entered in a software system, instituted in 1991, that checked for internal consistency. At approximately one fifth of hospitals, the data were entered by the computer system that year. At the other hospitals the data were entered manually. At the state level, the data entry system performed checks with both electronically and manually entered data. For example, gestational age and birth weight were checked for gross inconsistencies.
Birth certificate data from 195,919 births were reviewed. Births resulting in infants who weighed under 500 g or were born before 24 weeks were presumed nonviable and therefore excluded (n = 649). Births in hospitals with fewer than 100 annual deliveries were also excluded (n = 466), because those were predominately emergency deliveries at hospitals without routine obstetric services. The outcome of interest was primary cesarean deliveries, so women with prior cesarean delivery were excluded (n = 21,624). The frequency of missing data ranged from 0.02% for maternal age to 2.5% for medical risks. Records with missing data for variables of interest were excluded (n = 12,427). After those exclusions, 160,753 births and 154 hospitals remained in the database. The total number of records in the database after exclusions stayed the same.
We examined potential independent variables that seemed clinically relevant and were known to the provider before delivery. For example, birth weight is often used as a surrogate for gestational age because the two are closely correlated and the latter might be less accurate. We analyzed birth weight and clinical estimate of gestational age and found that both variables gave almost identical results. We chose to use clinical estimate of gestational age because birth weight was known only after delivery.
Certain variables from the database were recoded to create clinically meaningful categories. Maternal education was grouped as greater than, equal to, or less than high school; prenatal care was recoded as the trimester in which care began; and gestational age was regrouped to reflect very preterm (before 30 weeks), preterm (30–36 weeks), term (37–40 weeks), and postdate (at or after 41 weeks) pregnancies. Maternal age was coded as 10–19, 20–34, or 35 or more years old. A new variable, intrapartum complications, was created to represent any of the following indications for cesarean delivery: abruption, placenta previa, breech or other malpresentations, or cord prolapse. A woman was determined to be nulliparous if the variables livebirth-now-dead and live-birth-now-alive were both coded as 0.
For other variables used, we retained existing coding from the birth certificate database. A medical risk factor was considered present if a woman had one or more of following conditions: anemia (hemoglobin less than 10 g/dL or hematocrit lower than 30% during pregnancy), cardiac disease, lung disease, diabetes, renal disease, genital herpes, hemoglobinopathy, hypertension (chronic or pregnancy-induced), eclampsia, incompetent cervix, hydramnios or oligohydramnios, previous infant over 4000 g, previous preterm or small-for-gestational-age infants, Rh sensitization, or uterine bleeding. Congenital anomalies included anencephalus, spina bifida or meningocele, hydrocephalus, microcephalus, central nervous system anomalies, heart malformations, circulatory or respiratory anomalies, rectal atresia or stenosis, tracheoesophageal fistula or atresia, omphalocele or gastroschisis, gastrointestinal anomalies, malformed genitals, renal agenesis, other urogenital anomalies, cleft lip or palate, polydactyly or syndactyly or adactyly, club foot, diaphragmatic hernia, other musculoskeletal or intugumental anomalies, Down syndrome, or other chromosomal anomalies. Self-reported race was categorized as white, black, or Asian. In Illinois, Hispanic ethnicity, also self-reported, is coded in addition to race. Thus, Hispanic ethnicity was considered a separate variable.
The data were analyzed using SAS statistical software version 6.12 (SAS Institute Inc, Cary, NC) to calculate the odds of cesarean delivery associated with all potential independent variables. Stepwise logistic regression was done on a randomly chosen 25% of the sample to develop a main effects model predictive of cesarean delivery, with P < .05 required for entry of a variable into the model.
The model was validated on the remaining 75% of the sample. Using a formula developed from the regression coefficients,7 each woman's probability of cesarean delivery was calculated. Then women were grouped into quartiles based on their individual probabilities of cesarean delivery. For each quartile, the predicted rate was determined as the average probability among all in that quartile. That predicted rate was compared to the actual cesarean delivery rate for those women.
The validated model was applied to compare performance across hospitals. Each hospital's predicted primary cesarean delivery rate was calculated as the average probability for all women who delivered at that hospital. Actual rates were then compared with predicted rates.
The unadjusted risk of primary cesarean delivery according to selected medical factors for the entire study population is shown in Table 1. Nulliparity, multiple gestation, and intrapartum complications (placental previa, abruption, cord prolapse, breech, or malpresentation) were the strongest predictors of cesarean delivery. Very preterm, preterm, and postdate deliveries, and medical risks and congenital anomalies were associated with modestly increased risk.
The unadjusted risk of cesarean delivery according to social and demographic factors is shown in Table 2. Black race was protective against cesarean delivery, as was Hispanic ethnicity, age under 20 years, and little or no prenatal care. Age at least 35 years and married status were associated with a modest increase in risk of primary cesarean delivery. Compared with women with a high school education, women with less than a high school education had a decreased risk of primary cesarean delivery and women with greater than a high school education had an increased risk of primary cesarean delivery.
Logistic regression analysis was done using the variables in Tables 1 and 2. The variables included in the final model, which was generated from a random sample of 25% of the population, are shown in Table 3. All variables were retained in the final model except white and Asian races, congenital anomalies, prenatal care starting in the second or third trimester, and mother's marital status. Adjusted odds ratios were of similar magnitude to unadjusted risks. After all other variables in the model were controlled for, Hispanic ethnicity and black race were no longer protective against cesarean delivery, but instead incurred increased risk. Maternal education beyond high school protected against cesarean delivery in the final model, whereas that factor incurred increased risk in the unadjusted analysis.
Validation of the model is shown in Table 4. For all four quartiles, the actual rates were within predicted confidence intervals (CIs). The number of women in each quartile varied because some had identical probabilities of cesarean delivery. The validated model was then used to generate an expected primary cesarean delivery rate for each hospital based on the patient mix for that hospital. Predicted hospital primary cesarean delivery rates ranged from 9% to 24%. Actual hospital rates ranged from 6% to 30%. Of the 154 hospitals in the study, 35 (23%) had actual primary cesarean delivery rates that were above the upper limit of predicted CIs, 89 (58%) had rates that were within predicted CIs, and 30 (20%) had rates that were below the lower limit of predicted CIs.
To show the comparison between actual and predicted rates, hospitals were divided into quartiles based on their actual rates. In the first quartile were hospitals with the lowest actual primary cesarean rates in the state, and in the fourth quartile were hospitals with the highest actual primary cesarean rates in the state. There were 38 hospitals in the first quartile, 40 in the second, 39 in the third, and 37 in the fourth. Figure 1 shows the actual primary cesarean rates, divided by quartile, and grouped by whether the hospitals in that quartile had actual rates that were above, within, or below predicted 95% CIs. The black bars show hospitals with actual primary cesarean delivery rates greater than the upper limit of predicted CIs. Those hospitals were found in all but the lowest quartile of actual rates. Eight of 35 hospitals (23%) whose actual rates were greater than predicted CIs were not in the top quartile of raw rates. In the highest quartile of actual rates, the striped bar shows that ten of 37 hospitals in the 4th quartile (27%) were within predicted CIs and were doing numbers of cesarean deliveries appropriate for the risk status of the population (Figure 1).
For the 35 hospitals with actual rates greater than the predicted CI, the difference between actual and predicted rates ranged from 2.3%–16.9%. Twenty-two of the 38 hospitals had actual rates higher by at least 5%. For the 30 hospitals that had actual rate less than the predicted CIs, the difference between actual and predicted rates ranged from 1.6%–6.5%. Three of 30 hospitals had rates that were lower by at least 5%.
The total cesarean delivery rate in the United States decreased to 20.8% in 1995 from a peak of 24.7% in 1988.8 Although a further decrease is considered desirable, it is not known what the ideal target rate should be. An appropriate standard for cesarean rate based on outcomes research has not been established, and to judge performance on the basis of an arbitrary rate target for all hospitals does not account for differences in population risk. We showed that primary cesarean delivery rates can be readily risk adjusted by using a state birth certificate database, and that our approach substantially changes how hospital performance is judged. Specifically, we found that about one fourth of Illinois hospitals with the highest unadjusted rates were delivering risk-appropriate care, and about one fourth of hospitals with inappropriately high rates for case mix would have been missed if only those in the highest quartile of unadjusted rates were targeted.
Two general categories of factors might explain the variation in primary cesarean delivery rates between hospitals—case mix and hospital performance. If case mix alone explained all variation, our methodology would have shown actual rates to be the same as predicted rates. As we anticipated, that was not the case, suggesting that hospital performance also influenced the variation in actual rates. We examined several variables that represented the effect of case mix on primary cesarean delivery rates, including some that have not been previously well described, such as inadequate prenatal care. We did not address the various factors that might be considered in the category of hospital performance. The potential effect of those factors, from physician practice style to hospital resources, has been documented.3,4,9,10 The extent to which those factors affect variation in cesarean delivery rates after adjusting for population risk has not been studied, based on our MEDLINE search from 1966 to October 1998 using the search terms “cesarean section,” “cesarean section repeat,” “vaginal birth after cesarean,” “cesarean,” “risk assessment,” “health services research,” “outcome assessment,” and “risk adjustment.”
Risk adjustment has been used in other areas of medicine,11–13 but is relatively new in obstetrics and gynecology. Elliot et al14 did early work in the field of risk adjusting cesarean rates in low-risk women by using stratified analysis. Lieberman et al15 recently reported adjustment of cesarean delivery rates between types of providers within a single hospital, and her methodology might be applicable on a larger scale. There has been little interest, however, in risk adjustment of cesarean delivery rates across entire populations. Keeler et al6 recently did a population-based study in the state of Washington. They combined birth certificate data and maternal and infant discharge data and used four validated models for prediction, one each for prior cesarean, breech, first birth, and other. Using those models to adjust for case mix, they predicted a cesarean delivery rate for all 80 hospitals doing deliveries in Washington. That study included repeat cesarean deliveries and excluded low-birth-weight and multiple gestations, and the investigators found that risk adjustment did not greatly alter hospital ranking. Using similar methods, we came to a different conclusion. We speculate that the difference in findings might be attributable to their inclusion of women with prior cesarean deliveries because the overall effect of risk adjusting primary cesareans rates can be attenuated by the inclusion of risk-adjusted repeat cesarean rates.
A recent study by Aron et al16 used a methodology very similar to ours; they examined 21 hospitals in the Cleveland area, abstracting data from hospital charts. Using a validated model to adjust for primary cesarean rates, they found a wide range of predicted rates across hospitals. When actual rates were compared with predicted rates, they found that 24% of hospitals changed outlier status. One of the strengths of that study was the high quality of data obtained from chart abstraction.
Our use of birth certificate data to develop a model for risk adjusting of cesarean delivery rates has the advantage of being complete across the population for the factors reported, and a well-organized system of collection is already in place in the United States. Items reported in each state are similar. Thus, our methodology could be easily applied to national data, allowing regional variations to be studied without expending the substantial effort of extracting data from medical records. There are disadvantages to birth certificate data. A birth certificate database lacks information on some potential predictors of cesarean delivery, such as management of labor and payer status.3,9,10 Some, but not all, of that information could be obtained by linking with other databases. Other items are not reliably reported.
Of particular concern, a recent study of birth certificates in Georgia documented suboptimal accuracy in coding the presence or absence of prior cesarean deliveries.17 Women who had two vaginal deliveries or two cesarean deliveries were classified correctly 99.4% and 96.9%, respectively. Women who had a vaginal delivery after a cesarean, however, were classified correctly only 42% of the time, and women who had a cesarean after a vaginal delivery were classified correctly 79% of the time. It is unclear how these types of misclassifications might have affected our results. However, the results of Aron et al,16 who used a database with direct quality control by the investigators, were almost identical to ours. That helps mitigate concern about potential flaws in using birth certificate data for our purpose.
We used a method that calculates a CI for each predicted rate. The limitation of that approach was that hospitals with small numbers of deliveries will have calculated CIs that are broad. Thus, some small hospitals with suboptimal performance will have actual primary cesarean delivery rates that still fall within predicted CIs. The greater the number of deliveries in a hospital, the narrower the CI for predicted rate and the more likely it is to be identified as an outlier. The relatively large discrepancy between actual and predicted rates found for most outliers suggested that this approach is not too stringent. Even so, comparing a hospital's predicted and actual rates over several years might be a better measure of performance than examining any one year because by chance alone 5% of hospitals would be expected to have rates that fall outside predicted CIs.
Our interest in this investigation was prediction, not causality. Accordingly, we did not pursue analyses that might have allowed insight into why the various factors identified were associated with increased or decreased risk of cesarean delivery. Some of the factors might have functioned as surrogates for unmeasured risks, such as illicit drug use. The large effects found from nulliparity, multiple gestation, very preterm birth, and diagnoses such as placenta previa and malpresentation, were readily anticipated. To the extent that the rates of those factors might vary substantially in different hospital populations, it is illogical to compare cesarean delivery rates without first taking such variation into account. Although risk adjusting by using average performance as a benchmark is not necessarily a true measure of quality, we believe that approach should be the first step in understanding variations in primary cesarean delivery rates.
1. Ventura SJ, Martin JA, Mathews TJ, Clarke SC. Advance report of final natality statistics, 1994. Monthly vital statistics report. Hyattsville, MD: National Center for Health Statistics 1996;44 Suppl:71.
2. McKenzie L, Stephenson PA. Variation in cesarean section rates among hospitals in Washington state. Am J Public Health 1993;83:1109–12.
3. Goyert GL, Bottoms SF, Treadwell MC, Nehra PC. The physician factor in cesarean birth rates. N Engl J Med 1989;320:706–9.
4. Stafford RS. The impact of nonclinical factors on repeat cesarean section. JAMA 1991;265:59–63.
5. Public Health Service. Healthy people 2000: National health promotion and disease prevention objectives. DHHS publication no (PHS) 91-50212. Washington, DC: US Department of Health and Human Services, 1991.
6. Keeler EB, Park RE, Bell RM, Gifford DS, Keesey J. Adjusting cesarean delivery rates for case mix. Health Serv Res 1997;32:511–28.
7. Health Care Financing Administration. Medicare hospital mortality information 1987, Washington, DC: US Department of Health and Human Services, 1988.
8. Curtin SC, Kozak LJ. Cesarean delivery rates in 1995 continue to decline in the United States. Birth 1997;24:194–6.
9. Tussing AD, Wojtowycz MA. The cesarean decision in New York state, 1986. Economic and noneconomic aspects. Med Care 1992; 30:529–40.
10. Localio AR, Lawthers AG, Bengtson JM, Hebert LE, Weaver SL, Brennan TA, et al. Relationship between malpractice claims and cesarean delivery. JAMA 1993;269:366–73.
11. Rosenthal GE, Harper DL, Quinn LM, Cooper GS. Severity-adjusted mortality and length of stay in teaching hospitals: Results of a regional study. JAMA 1997;278:485–90.
12. Iezzoni LI, Schwartz M, Restuccia J. The role of severity information in health policy debates: A survey of state and regional concerns. Inquiry 1991;28:117–28.
13. Leung KM, Elashoff RM, Rees KS, Hasan MM, Legorreta AP. Hospital- and patient-related characteristics determining maternity length of stay: A hierarchical linear model approach. Am J Public Health 1998;88:377–81.
14. Elliot JP, Russel MM, Dickason LA. The labor-adjusted cesarean section rate. A more informative method than the cesarean section “rate” for assessing a practitioner's labor and delivery skill. Am J Obstet Gynecol 1997;177:139–43.
15. Lieberman E, Lang JM, Heffner LJ, Cohen A. Assessing the role of case mix in cesarean delivery rates. Obstet Gynecol 1998;92:1–7.
16. Aron DC, Harper DL, Shepardson LB, Rosenthal GE. Impact of risk-adjusting cesarean delivery rates when reporting hospital performance. JAMA 1998;279:1968–72.
17. Green DC, Moore JM, Adams MM, Berg CJ, Wilcox LS, McCarthy BJ. Are we underestimating rates of vaginal birth after previous cesarean birth? The validity of delivery methods from birth certificates. Am J Epidemiol 1998;147:581–6.