Comparing the frequentist and Bayesian estimation approaches for cox regression in modelling cervical cancer patient outcomes in Uganda.

Abstract

The objective of this paper was to compare the Frequentist and Bayesian estimation approaches for Cox regression in modelling cervical cancer patient outcomes in Uganda by evaluating the predictive performance of the two approaches, and assessing the association of clinical and sociodemographic factors with patient survival. The Cox PH, Exponential PH, and Weibull PH models were run under both frameworks, and their model performance and predictive accuracy was compared to determine which framework best fits survival analysis for cervical cancer patients in Uganda. The models were applied on dataset containing 454 women diagnosed with Cervical Cancer which was primarily collected by the Uganda Cancer Institute (UCI) between June 2019 to February 2020. The study also observed how different covariates such as age at diagnosis, morphology, initial disease stage, region of treatment, and other clinical and sociodemographic factors influence the time until death for the cervical cancer patients involved in the study. The non-parametric Cox PH model, and the Exponential PH model demonstrated the best overall performance based on model fit statistics under the Frequentist and Bayesian approaches respectively. Both models yielded nearly identical predictive outcomes when evaluated against each other with C-index ≈ 0.8 and AUC ≈ 0.79. Further findings based on the Bayesian Exponential model revealed that, diagnosis at stage 3 (HR=5.259; CrI: 1.446, 25.798) and stage 4 (HR=6.678; CrI: 1.63, 35.196), treatment by palliation (HR=4.940; CrI: 1.473, 20.503), and unspecified morphology (HR=2.174; CrI: 1.231, 3.785), were significantly associated with lower chances of survival. While women with negative metastases (HR=0.505; CrI: 0.284, 0.944), and those who were treated from the Eastern (HR=0.115; CrI: 0.029, 0.36) and Central (HR=0.498; CrI: 0.249, 0.985) regions, exhibited reduced hazard of death. Nevertheless, age at diagnosis, occupation and marital status had no significant effect on the survival outcomes of the women with involved in this study. In conclusion, the study underscores the critical role of stage at diagnosis in determining cervical cancer survival, it therefore calls upon the Ministry of Health to strengthen routine screening initiatives to promote early detection. The findings of the study also showed that both the Frequentist and Bayesian approaches perform effectively and can be applied in survival analysis. The study therefore recommends that future research should evaluate Frequentist and Bayesian frailty survival models to account for unobserved heterogeneity among cervical cancer patients.