Risk adjustment for cesarean delivery rates: how many variables do we need? An observational study using administrative databases

Background Various studies indicate that inter-hospital comparisons have to take case mix into account and that risk adjustment procedures are necessary to control for potential predictors of cesarean delivery (CD). Different data sources have been used to retrieve information on potential predictors of CD. The aim of this study was to compare the discrimination capacity and fit of predictive models of CD created using different sources and to assess whether more complex models improve inter-hospital comparisons. Methods We created 4 predictive models of CD. One model included only variables from Hospital Discharge Records of the index hospitalization, one included also information from previous hospitalizations, one also clinical variables from birth certificates (BC) and one also socio-demographic variables. We compared the four models using the Receiver Operator Curve and the Akaike and Bayesian Information Criteria. Results Information from Birth Certificates improved the discrimination and model fit. Adding socio-demographic variables or past comorbidities did not improve the discrimination capacity or the model fit. Hospital-specific CD resulting from the models were highly correlated. Conclusions Record linkage improves the performance of the models but does not affect inter-hospital comparisons.