Purpose: To assess whether the addition of splines and interaction terms improve the performance of a validated risk-adjustment model for primary cesarean rates.

Methods: We built several logistic regression models for primary cesarean delivery using maternal risk factors.  We utilized California birth certificate data for 2003 linked to hospital discharge data for mothers and babies.  Model predictors included: maternal age, race, and medical conditions, gestational age, multiple births, insurance, nulliparity, complications of pregnancy, and the trimester in which prenatal care began.  Our initial model (validated for other states) uses linear functions of main effects for all independent variables.  Here, we augmented this using subgrouping, smoothing and cubic spline function techniques to capture potential non-linearity and interaction effects. We report bootstrap estimates of model calibration and discrimination through C statistics (area under the receiver operating characteristic curve), Brier scores, and calibration plots.

Results: After cleaning, our models describe 382,566 births. We report on three logistic regressions motivated by exploratory analyses.  Model A incorporates key predictors as linear main effects (C statistic .765). Model B uses restricted cubic splines to capture potential non-linearity in maternal age (C statistic .766).  Model C adds interactions of maternal age with other key predictors (C statistic .766).  Brier score was 0.117 for each.                    

Conclusions: Despite significant differences in likelihood ratio testing (p<.0001), all three models show comparable discrimination and calibration, suggesting no real performance advantage for the more complex models. In the interests of parsimony and improving clinical understanding, we plan to focus future work on model A, which mirrors our previously validated approaches.