Sensitivity Analysis of the Quadratic Discriminant Function for Predicting Pregnancy Outcomes

Abstract

The prevalence rate of stillbirth is ten times higher in developing countries relative to developed countries with a 2016 rate of 18 percent in Ghana. This study employed the Quadratic Discriminant Function for discriminating and classifying of pregnancy outcomes based on some predictors. The study further examined the sensitivity of the Quadratic Discriminant Function in predicting pregnancy outcomes with variations in the training and test samples of deliveries recorded in a hospital in Accra, Ghana. The study considered the scenarios; 50:50, 60:40, 70:30 and 75:25 ratios of training sets to testing sets. Predictor variables on both maternal factors (maternal age, parity and gravida) and fetus variables (weight at birth and gestational period) were all statistically significant (P < .01) in discriminating between live birth and stillbirth. Results showed that maternal age had a negative effect on the live birth outcomes, while parity, gravida, gestational period and fetus weight recorded positive effects on live birth outcomes. The 75:25 ratio outperformed the other ratios in discriminating between live and stillbirth based on the Actual Error Rate of 7.28% compared to 7.81%, 12.14% and 13.79% for the 50:50, 70:30 and 60:40 ratios respectively whereas, the receiver operating characteristic curve shows the 70:30 (AUC= 0.9233) ratio outperformed the others. The study recommend the use of either the 70:30 or 75:25 training to test ratios for classification and discrimination related problems. Moreover, further research to establish the power of the respective training to test sample ratios with other statistical classification tools and more socio-economic variables can be considered.