Comparison of models for analysis of time to loss to follow up and associated risk factors among patients receiving tuberculosis treatment from Mbarara Regional Referral Hospital, Uganda

Abstract

Background: Tuberculosis (TB) remains the top infectious killer worldwide. In Uganda, incidence of TB was reported at 200 per 100,000 people per year. Loss to follow up (LTFU) on TB treatment was 13% in Uganda and 10% in Mbarara according to the national TB and leprosy report of 2018. The proportion of patients who successfully complete treatment is one of the priority indicators of the TB program performance. A better understanding of the time when patients get LTFU and assessment of the associated risk factors are cornerstones for designing time relevant intervention strategies. Methodology: The outcome variable was time to LTFU, recorded in weeks. The source of the data was the National Tuberculosis and Leprosy Programme Unit TB registers for the years 2017 to 2020. Kaplan Meier method was used to obtain univariate descriptive statistics and compare the survival experience by the categorical variables. Bivariate and multivariable analyses were performed on both the Cox and the parametric models to obtain factors associated with time to LTFU and then the models compared on the model fit using Akaike’s Information Criteria (AIC). The frailty model was used to measure the covariate effects in order to describe the influence of unobserved covariates in the model. Results: The eligible TB patients for the study were 917. The incidence rate of LTFU on TB treatment was 17.6 cases per 100 PYO whereas the median time to LTFU was not defined. The results from the AIC indicated that the parametric models outperformed the Cox PH model. The lognormal distribution was the best model with AIC 741.25 whereas the Cox model had AIC 989.00 The results from the multivariable analysis indicated that HIV status and TB class were the only significant variables for the Lognormal model. Patients who were HIV positive were lost to follow up at least 3 times later than the HIV negative patients and those who had the drug sensitive TB were LTFU at least 9 fold later than the drug resistant patients holding other factors constant (TR=3.33, P-val=0.02, 95%CI=1.23-9.00) and (TR=9.21, P-val=0.02, 95%CI=1.44-58.71) respectively. There was no evidence of heterogeneity among the patients since the p-value of the frailty model was not statistically significant. Conclusion: Since TB is still a threat in Uganda, the TB patients at the hospital should be informed about the consequences of interrupting treatment the moment they get absorbed in the care The parametric models outperformed the Cox PH model in the model fit thus researchers should always consider analysis using the parametric models if the conditions underlying the models are met since they provide more accurate estimates