Since 1995 in the Emilia Romagna Region of Northern Italy (RER), all hospital discharge abstracts have been electronically recorded, using a Hospital Information System (HIS). The data stored in the system includes demographics [ID number, gender, date and place of birth, place of residence], discharge ID, admission and discharge dates, up to 9 discharge diagnoses and 9 procedures (International Classification of Diseases, 9thRevision, Clinical Modification ICD-IX-CM), ward(s) of hospitalization, date(s) of in-hospital transfer, and the regional code of the admitting facility.
We selected all hospital discharge abstracts for women in labor and of newborns from 36 maternity units in the region from January 2003 to December 2004.
This study takes as its sample live births for whom the discharge records for the mothers and infants were linked by hospital code, mother's discharge ID and date of delivery.
To identify the delivery, we used Diagnosis-Related Groups (DRGs) 370–375 from the discharge data. DRG 370 and 371 (cesarean section with and without complication, respectively) were used to identify cesarean deliveries. ICD-IX-CM diagnosis code 654.2x was used to identify any previous cesarean deliveries . The number of births from primary cesareans was calculated as the difference between the number of births from c-sections deliveries and number of births from c-sections deliveries in women with previous cesareans.
Therefore, primary cesarean rates were calculated with the formula:
There were 62,836 births from deliveries to women with no previous c-section and they were included in our study population, excluding the following:
mothers under 11 and over 50 years of age
mothers discharged from hospitals without an operating room
infants with a birth weight under 550 or over 6000 g
Hospitals with fewer than 100 deliveries per year were excluded to warrant sufficient power of comparison.
The following socio-demographic variables, considered as potential risk factors for cesarean sections, were collected: maternal age (<17, 18–20, 21–24, 25–28, 29–33, 34–38 ≥ 39), citizenship, (Italian, from developing countries, undeveloping countries other than Italy), residency (RER or other), and marital status (married, divorced-separated, single, widow). Maternal and neonatal clinical factors were also retrieved. These factors were defined using the primary and secondary discharge diagnoses of the delivery and newborn admission ' [see Additional file 1]'.
We did not consider dystocia and fetal distress as potential risk factors because of the poor reliability of their definition and because this diagnosis may reflect post ad hoc justifications of cesarean use, rather than objectively assessed conditions [17, 18].
The study was conducted in collaboration with the Azienda Ospedaliera Sant'Orsola-Malpighi, the teaching hospital of the University of Bologna, Italy
Descriptive statistics and hospital-specific crude Odds Ratios (odds of c-section for patients admitted to a specific hospital vs. odds of c-section for patients admitted to the reference category) were reported.
To take into account the role of confounders, two different logistic regression models were adopted: a "full" and a "parsimonious" model.
The "full" model was defined applying a backward selection procedure to a list of potential confounders selected according to available scientific evidence. All previously defined factors were entered and were retained if they were significant predictors of c-section (p < .05). Because of the large size of the database, an α of .05 was chosen to minimize the number of variables in the model and to maximize the strength of the association.
The "parsimonious" model was defined applying a "change-in estimate" procedure [19–21]. The first step of this method included the same factors entered in the full model and the exposure of interest (a specific hospital vs. reference category). Subsequently, all factors that did not modify, or only slightly modified the estimated effect of exposure, were excluded from the model.
The "change-in estimate" procedure identified the actual confounders for single comparisons and was repeated for each comparison (each hospital vs reference), defining as many risk adjustment models as there were comparisons. All factors, identified by at least one comparison, were included in the "parsimonious" model.
The model's performance was evaluated based on how closely it predicted the results actually observed, following the criteria for discrimination (C statistics) and calibration (Hosmer-Lemeshow test). The differences in the predictive value of the two models were assessed using the Akaike Information Criterion  to augment the log likelihood ratio χ2 test, with a penalty for differences in the number of variables in the models compared.
The reference category included hospitals with the lowest adjusted c-section rates based on the full model. This category was defined according to the following steps:
28 hospital dummies were added to the full model and the corresponding adjusted ORs were ranked. In this case the reference category was selected as the hospital with the highest number of births.
Four hospitals with the lowest adjusted ORs were selected as reference category.
Finally 25 hospital dummies, representing the rest of the 25 hospitals, were added to the full model and the estimated ORs were ranked. In this case the four hospitals, selected as reference category, were used for benchmark purposes in evaluating hospital performance for c-section in this study.
The crude and adjusted ORs obtained by the two models were used to rank hospitals, and the consistency of rankings was assessed using Spearman's rank correlation coefficient.
The statistical analysis was performed using SAS 8.2 (SAS Institute, Cary, NC) and Stata 8.2. (College Station, Texas 77845, USA).