Multiple imputation using chained equations for missing data in survival models: applied to multidrug-resistant tuberculosis and HIV data

Sizwe Vincent Mbona; Principal Ndlovu; Henry Mwambi; Shaun Ramroop

doi:10.4081/jphia.2023.2388

Multiple imputation using chained equations for missing data in survival models: applied to multidrug-resistant tuberculosis and HIV data

J Public Health Afr. 2023 Jun 5;14(8):2388. doi: 10.4081/jphia.2023.2388. eCollection 2023 Aug 7.

Authors

Sizwe Vincent Mbona¹, Principal Ndlovu², Henry Mwambi³, Shaun Ramroop³

Affiliations

¹ Department of Statistics, Durban University of Technology, Durban.
² Department of Statistics, University of South Africa, Pretoria.
³ School of Mathematics, Statistics and Computer Science, University of KwaZulu-Natal, Pietermaritzburg, South Africa.

Abstract

Background: Missing data are a prevalent problem in almost all types of data analyses, such as survival data analysis.

Objective: To evaluate the performance of multivariable imputation via chained equations in determining the factors that affect the survival of multidrug-resistant-tuberculosis (MDR-TB) and HIV-coinfected patients in KwaZulu-Natal.

Materials and methods: Secondary data from 1542 multidrug-resistant tuberculosis patients were used in this study. First, data from patients with some missing observations were deleted from the original data set to obtain the complete case (CC) data set. Second, missing observations in the original data set were imputed 15 times to obtain complete data sets using a multivariable imputation case (MIC). The Cox regression model was fitted to both the CC and MIC data, and the results were compared using the model goodness of fit criteria [likelihood ratio tests, Akaike information criterion (AIC), and Bayesian Information Criterion (BIC)].

Results: The Cox regression model fitted the MIC data set better (likelihood ratio test statistic =76.88 on 10 df with P<0.01, AIC =1040.90, and BIC =1099.65) than the CC data set (likelihood ratio test statistic =42.68 on 10 df with P<0.01, AIC =1186.05 and BIC =1228.47). Variables that were insignificant when the model was fitted to the CC data set became significant when the model was fitted to the MIC data set.

Conclusion: Correcting missing data using multiple imputation techniques for the MDR-TB problem is recommended. This approach led to better estimates and more power in the model.

Keywords: missing data; multidrug-resistance tuberculosis; multiple imputation.

Grants and funding

Funding: none.