Abstract

Studies of neonatal mortality determinants often rely on retrospective data since large-scale follow-up data are generally not available. A characteristic of neonatal data collected retrospectively from cross-sectional studies is that births and deaths are recorded within a fixed interval that is long compared to the risk period (first 28 days of life). Furthermore, a substantial proportion of subjects never die due to neonatal causes and survive past the risk period. These subjects are cured and therefore complete observations. The Cox proportional hazards (Cox PH) regression model has been a popular tool in the study of causes of neonatal death. It assumes that given a long enough follow-up time, all observations will eventually experience the event of interest. This assumption may not hold for retrospective neonatal mortality data. Thus, the use of standard Cox PH model that treats cured observations as censored may not be appropriate. In such a case, the mixture cure rate model (MCRM), which models the cure rate (incidence part) and survival time (latency part) separately may be a better alternative. The objective of this study is twofold: (i) to determine under what conditions the Proportional Hazards MCRM is preferable to Cox PH Model (ii) to reanalyze the Bangladesh Demographic and Health Survey (BDHS) 2017 data for causes of neonatal mortality using MCRM and compare the results with traditional Cox PH. To realize the first objective, the models are compared (with respect to bias and standard error of covariate estimates, coverage probability of confidence interval and power of the test of no covariate effect) using censored survival data simulated under six different settings involving the censoring proportion and cure fraction. Simulation experiments revealed better performance of MCRM when the cure fraction was nearly as high as the censoring proportion – a condition observed in the BDHS data. Analysis of the BDHS data using the Cox PH model revealed only three variables, namely, mother’s education in years, type of birth of child and place of delivery to be significant determinants of neonatal survival. On the hand, analysis via MCRM indicated that mother’s education, mother’s age, type of birth, number of antenatal care (ANC) visits and type of delivery assistant had significant effects on the probability of neonatal death (incidence) while the variables exposure to mass media, mother’s age, length of previous birth interval, number of ANC visits, type of delivery and whether baby was checked within 24 hours had significant effects on time to neonatal death among the uncured (latency). Thus, Cox PH failed to capture the effects of ANC, delivery and postnatal care-related variables that are deemed important by the literature. In conclusion, by measuring both short and long-term effects of covariates on neonatal survival, the MCRM gives better insights into the mechanisms through which covariates affect the outcome compared to the Cox PH model and should therefore be the model of choice when analyzing neonatal mortality data with high censoring proportions.