Evaluation of survival models in examining factors associated with antenatal visits initiation among pregnant women in Nebbi district, Uganda
MetadataShow full item record
Introduction and background: The timing of antenatal (ANC) visits initiation is paramount in ensuring optimal maternal health outcomes. Both parametric and semi-parametric survival analysis models have been used to identify the predictors of ANC visit initiation. The semi-Parametric Cox model has been the most commonly used because of its distribution-free assumption and yet there are times when the assumption of constant proportional hazard may be violated. The violation of the proportional hazard assumption renders the application of semi-parametric Cox model inappropriate. Comparison of the semi-parametric Cox model with the parametric has always shown that the latter outperform the former. Problem: Several studies on factors associated with antenatal visits initiation using survival analysis models have been conducted. A number of studies have evaluated parametric and semi-parametric survival analysis models using Akaike Information Criterion (AIC) and yet AIC does not provide a test of model quality in its absolute sense. Few studies have evaluated parametric accelerated failure time and semi-parametric Cox models using institutional-based antenatal data based on metrics of model quality measurement such as receiver operating characteristic (ROC) curves and brier score index (BSI). Objectives: This study aimed at evaluating parametric Accelerated Failure Time (AFT) and semi-parametric Cox models in examining the predictors of ANC visits initiation among pregnant women in Nebbi district, Uganda Method: A retrospective, hospital-based HMIS data from the medical records of 1,688 pregnant women for the period of July 2018 to June 2019 were studied. Both semi-parametric Cox and parametric Accelerated failure time AFT) models were used to analyze the effects of the covariates on the ANC visits initiation. Kaplan-Meier curves were constructed to compare the survival functions of the sub-groups. Brier Score index, receiver operating Characteristic (ROC) curves and its Area under the curve AUC), Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) were statistical metrics used to compare the goodness of fit of the two classes of models. More so, Cox-Snell’s residual and martingale residual plots were used to compare the model goodness of fit by use of residual plots. All the analysis was done in STATA ver15. Results: The results from the multivariable semi-parametric Cox and parametric AFT models analysis show similar effects of the covariates on ANC visits initiation. The results indicate that the parametric AFT models have better goodness of fit for the data compared to semi-parametric Cox model. Among the parametric AFT class of models, the log-normal indicated the best model fit for the data as portrayed by the different statistical metrics of comparison used. The predictor variables of male involvement and gravida were statistically significant across all the classes of parametric AFT and semi-parametric Cox models (p-value <0.05). The results further show that pregnant women whose male partners were involved in care, those who reside in urban settings and those who were prime gravida initiate ANC visits earlier compared to those women whose partners were not involved and were multi-gravida i.e urban residence [(TR= 0.980, p-value =0.049, 95%CI=0.960 – 0.100)]; multi-gravida [(TR =1.040, p-value =0.008, 95%CI=1.011-1.071)] and male partner involved [(TR= 0.967, p-value = 0.009, 95%CI = 0.943-0.992)] based on the parametric AFT model of log-normal coefficients. The proportional hazard assumption was violated when tested using scaled Schoenfeld test. Conclusion: The results indicated that the parametric AFT model especially the log-normal model had the best model fit for the data compared to other models. The parametric AFT models especially the log-normal should be used as the most suitable modelling method for a right-censored data when the proportional hazard assumption is violated.