Comparison of semi-parametric and parametric survival analysis models for identifying predictors of virological suppression among HIV-infected adults on ART at the Joint Clinical Research Centre, Uganda

Abstract

Background: Semi-parametric and parametric survival analysis models have been widely used to analyze the effects of covariates on survival time. The Cox proportional hazards (PH) model in particular remains the most popular model used for survival modeling due to its minimal assumptions. However, its key assumption of proportional hazards may sometimes be violated. Suitability of both semi-parametric Cox PH and parametric models depends on the data used and the study population. If wrong model is applied to data, it gives inaccurate estimates and misleading inferences.

Objectives: The study objectives were: (i) To identify the best-fitting survival analysis model, semi-parametric or parametric, for identifying predictors of virological suppression among adult patients living with HIV/AIDS on ART at Joint Clinical Research Center (JCRC), Uganda; (ii)To identify predictors of virological suppression among HIV-infected adults on ART at JCRC, Uganda.

Methods: This study adopted a retrospective cohort design, and used secondary data of HIV-infected adults on ART which is routinely collected at JCRC, Uganda. The target population was adult HIV patients years’ age, enrolled between January 2016 and October 2020. Exploratory data analysis was carried out to generate descriptive statistics for the variables. Kaplan–Meier survival curves and long-rank test were used to describe survival experiences of categorical variables. Missing data was imputed using multiple imputation by chained equations (MICE). The semi-parametric Cox PH model and parametric accelerated failure time (AFT) models (Exponential, Weibull, Log-logistic, Lognormal and Gamma) were fitted, and the best-fitting model was selected basing on the concordance index (C-index). The selected model was then used for analysing predictors of virological suppression.

Results: The overall median time to virological suppression was 12(95% CI: 10-14) months. Concordance indices for log-logistic model were relatively higher and closer to 1 compared to other models. Baseline viral load (Time Ratio (TR) =1.080, P-value=0.020) was significantly associated with virological suppression. Conversely, older adults (TR=1.281, P-value=0.206),weight (TR=0.999, P-value=0.886), WHO clinical stages 2 (TR=0.909, P-value=0.495), 3 (TR=1.126, P value=0.506) and 4 (TR=0.738, P-value=0.299), female sex (TR=1.000, Pvalue=0.998), CD4 count ≥200 (TR=0.808, P-value=0.131), PI-based ART regimen (TR=0.772, P-value=0.367), and other regimens (TR=0.652, P-value=0.083) were not significantly associated with virological suppression.

Conclusion: The parametric Log-logistic model fitted the HIV/AIDS data of adult population better than other parametric AFT models and the semi-parametric Cox PH model. Baseline viral load was the only significant predictor of virological suppression among HIV-infected adults on ART. Although the Log-logistic model was the best fitting model, its use should take into account the study limitations. Rigorous comparison of performance of different survival analysis models is recommended when analysing predictors of virological suppression for HIV-infected adults on ART.