Data collection
All live births (n = 176,399) occurring in either hospitals or obstetrics/gynecology clinics between 1 January and 31 December 2003, and for whom birth certificate data could be linked with the NHI claims data, were selected as the sample for this study. The mother's date of birth, along with her unique personal identification number, provided the link between the birth certificate data and the NHI claims data.
The birth certificate dataset contains various parental demographics (including age, the highest education level achieved, marital status, and county of residence), infant gestational age (in weeks), birth weight (in grams) and gender, as well as details on multiple pregnancies and the mother's gravidity. The NHI claims data contain information on all deliveries occurring in NHI-contracted hospitals and clinics (over 92% of all healthcare institutions), including the method of delivery, the characteristics of the hospital/clinic and attending physicians, as well as one principal diagnosis code, and up to four secondary diagnosis codes for each hospital admission, from the International Classification of Disease, Ninth Revision, Clinical Modification (ICD-9-CM).
Those cases with missing data (n = 924), and those cases for which delivery was carried out by physicians who had fewer than 20 deliveries in 2003 (n = 2964), were excluded from the sample. The reason for the exclusion of the latter group of physicians was to ensure that all physicians included within this study had adequate obstetrics practical experience. We were ultimately left with a study sample of 172,511 deliveries for analysis in this study, comprising 60,079 births by CS and 112,432 vaginal deliveries.
The availability of unique physician identifiers within the claims data for each medical claim submitted enabled us to identify the same physician carrying out one or more deliveries in 2003; data for that year indicated that 1031 physicians delivered the babies of the 172,511 sampled patients. We divided the physicians into four equivalent groups based upon the quartile distribution of their crude (actual) CS rates during the period of study. Those physicians whose CS rates fell into the top quartile had the highest crude CS rates, while those falling into the bottom quartile had the lowest crude CS rates.
The institutional review board (IRB) at Taipei Medical University, Taipei, Taiwan granted ethical approval for the study.
Variable definitions
'Case mix' represents those factors demonstrated in the literature to increase the risk of CS delivery. In this study, we categorized these factors into obstetric, pregnancy, and other risk factors.
Obstetric risk factors
Numerous studies have documented a breech presentation, dystocia, and fetal distress as obstetric factors which are indications for a CS [5, 11–14]; however, while these are common indications for a CS, dystocia and fetal distress diagnoses are often very subjective, and, indeed, are often not risk factors in themselves. The inclusion of these factors in the regression, along with preexisting risk factors, may well have masked many important differences by 'adjusting away' subjective practice differences among physicians. Furthermore, since dystocia may be related to fetal macrosomia, and fetal distress may be related to various other conditions (including diabetes, hypertension, and collagen vascular disease), this may once again have introduced potential redundancies into the regression.
There is also a lack of standard clinical criteria for defining dystocia and fetal distress [8, 10], with obstetricians in different healthcare institutions possibly applying the terms to quite-different conditions; thus, the variability in the proportions of patients diagnosed with dystocia and fetal distress may be partially attributable to differences in defining these conditions, rather than differences in the physician/patient mix. We therefore selected malpresentation (ICD-9-CM codes 652, 761.7, 763.0, or 763.1), a prolapsed cord (663.0), antepartum hemorrhage, abruptio placenta, and previa placenta (641, 762.0 or 762.1) to better define and represent the obstetric risk factors likely to result in a CS.
Pregnancy risk factors
A number of independent factors, all of which seem clinically relevant and are known to physicians prior to delivery, were included in this study. A woman was considered to have a pregnancy risk factor if she had one or more of the following conditions: a previous CS history (654.2), genital herpes (647.6), diabetes mellitus (648.0, 648.8, or 775.0), anemia (648.2), cardiac disease (648.5 or 648.6), arterial hypertension (642.0, 642.1, 642.2, 642.3, 642.9, or 760.0), eclampsia/preeclampsia (642.4, 642.5, 642.6, or 642.7), polyhydramnios/oligohydramnios (657.0 or 658.0), infection of the amniotic cavity (658.4), a congenital/acquired abnormality of the cervix or vagina (654.6 or 654.7), iso-immunization with Rh antigen (656.1), insufficient or excessive fetal growth (656.5 or 656.6), cerebral occlusion-hemorrhage (430, 431, 432, 433, or 434), premature labor (644), a multiple gestation (651), premature rupture of the membrane (658.1), or cervical incompetence (654.5).
Other risk factors
Certain variables from the database were also recorded to create clinically meaningful categories. These variables included maternal age (in years) at the time of the infant's birth, infant gender, and parity. Maternal ages were grouped into < 20, 20–34 and ≥ 35 years. Since birthweight and gestational week are highly correlated, to prevent redundancy by including both parameters, gestational age was selected to capture the effect of preterm (< 37 weeks) or postdate delivery (≥ 42 weeks). Insufficient or excessive fetal growth (656.5 or 656.6) was also included as a pregnancy risk factor to further distinguish the risk of intrauterine growth retardation or macrosomia for a given gestational age. Parity was recorded as whether or not a mother was parous.
Statistical analysis
All of the statistical analyses within this study were performed using the SAS statistical package (SAS System for Windows, version 8.2, Cary, NC). The χ2 test was first performed as a means of assessing whether there were any significant variations in the distribution of patient population risks across the physician quartiles, and second, to evaluate whether there were any significant differences, again by physician quartiles, in the cesarean delivery rates among women for each of the risk factors.
A univariate analysis was first carried out as the primary method of calculating the odds of a cesarean delivery, in order to determine whether there was any association with the potential risk factors. Stepwise logistic regressions were then conducted on parturients delivered by physicians in all four quartiles (referred to as model 1) to first develop a predictive model, and second, as a means of minimizing the number of predictive risk factors within the formula. A p value of < 0.05 was required for entry of a risk factor into the model. Given the sufficiently large volume of sample patients adopted for the current study, it was extremely unlikely that any important variables would have been overlooked.
Having computed the predicted risk of CS for each woman based upon the predictive model, these predicted risks were then aggregated by physician to determine the expected (risk-adjusted) CS rate and 95% confidence interval (CI) for each physician. The actual rates were then compared with the expected CS rates for each physician so as to determine how many of the physicians were below, within, and above the predicted CI in each physician quartile.
The main problem with risk adjustment, however, is the way in which an appropriate 'gold standard' is applied. Deriving a logistic equation and then applying it back to the same population from which it was derived is a well-accepted technique; however, there is a certain circular logic to this method, since the gold standard is partially defined by practices that are themselves outliers, such as the case in model 1.
Since the average CS rates for physicians in quartile 4 were as high as 52%, we therefore treated physicians in quartile 4 as outliers and repeated the logistic regression using only those parturients delivered by physicians in quartiles 1, 2, and 3. The results of this analysis are referred to as model 2. Based upon the regression results from model 2, we then recalculated the predicted CS rate and the 95% CI for each physician in the four quartiles, and again determined how many of the physicians were below, within, and above the predicted CI in each physician quartile. Our intuition was that there would be even-greater variation in predicted cesarean section rates under model 2 than under model 1.