Model-based Fully Conditional Specification and Predictive Mean Matching: Application to Hiv Risk Factors Among Female Sex Workers in Kenya

Abstract

Background: HIV disproportionately a ects sex workers. It is important to continually evaluate sex work, given it’s uid and dynamic nature. Missing data is a common complication to HIV research, especially where accurate and complete collection of data is a challenge. Aim: To study the missing data problem in the female sex workers’ data and employ the multiple imputation technique. Methods: Multiple imputation using the Fully Conditional Speci cation (FCS) was used to handle the missing data problem. For the target analysis, a binary logistic model was used to test association between HIV status and risk factors among female sex workers. We assessed the impact of missing data on the statistical signi cance of the risk factors of HIV. We further, compared the performance of model-based FCS and Predictive Mean Matching (PMM) by assessing distributional properties, convergence, adjusted odds ratios, interval width and relative e ciency. Results: There were generally low proportions of missingness and missing data was not found to a ect statistical signi cance of associations of HIV risk factors to HIV positivity of female sex workers. There was a reverse in the interpretation of results in the number of sex acts per week, though not statistically signi cant. Multiple imputation reduced standard errors of parameter estimates, giving more precise estimates and narrower con dence intervals. Distributional properties were also preserved by MI. Model-based FCS performed slightly better in convergence, interval width while PMM had better relative e ciency. Conclusion: Multiple imputation results in more reliable estimates with lower standard errors. Performance of the model-based FCS was considerably better than PMM. These results are, however, not considered conclusive and may need validation using a large simulation study.