Modelling The Risk Factors Of Neonatal Mortality Using Survival Analysis

ABSTRACT

Several strategies have been put in place in an attempt to reduce childhood mortality in

Ghana, however the proportions of death among neonates are still quite high. The study

therefore seeks to model neonatal mortality using survival analysis approach. The data used

for the study was obtained from the neonate’s folders at St. Jude Hospital in Obuasi in the

Ashanti Region between January 1, 2012 and December 31, 2015. Data on maternal

characteristics was also obtained. Neonates who were born before the 28th day and those who

have experienced the event (death) were considered for the study. The study employed the

Kaplan Meier (K-M) and Log rank test for the descriptive analysis.

The Cox PH and Parametric models (Exponential, Weibull, Gompertz, Log-logistic and Lognormal)

were fitted to the neonatal data and their results were compared using the AIC to

determine the best model to explain survival of neonates. A semi parametric shared frailty

model was also fitted to the data to examine whether there are unobserved heterogeneity

among neonates at the community level. The Proportional Hazards assumption was checked

using both graphical methods and the PH assumption test based on the Schoenfeld residual

and was observed that the PH assumption was not violated. Results from the study showed

that hazard ratios for the PH models (Cox, Exponential, Weibull and Gompertz) were similar,

however comparison of the PH models using the AIC showed that the Gompertz PH model

best fit the data.

A comparison of AFT models (Weibull, Exponential, Lognormal, Gompertz, and Log

logistic) also showed that the Lognormal AFT fit the data best. A comparison of the best PH

(Gompertz PH) and AFT (Lognormal AFT) model using the AIC showed that the Gompertz

PH is the best model in predicting neonatal survival. Parity, Apgar score 1, birth weight and

iii

place of residence were significantly related with neonatal mortality. A comparison of the

shared frailty models (Cox, Exponential, Weibull, Gompertz, Lognormal and Log-logistics)

using AIC revealed that exponential distribution with Gamma frailty is the best model for

checking the unobserved heterogeneity in the data. Unobserved heterogeneity in categories of

neonates based on place of residence was found.