Using advanced statistical methods to improve prediction of HIV outcomes from a routine observational clinical database: Validation of findings using a nested research cohort database

Abstract

Background Observational HIV clinical databases are increasingly being used to address questions related to HIV care and treatment. Advantages associated with these databases are the fact that they are of large sample size, and estimates obtained from them tend to be more generalizable, compared to results from specialized cohorts, randomized controlled trials and other study designs. However they are often characterised by inadequate data quality and several biases that need to be addressed prior to using them for research. Consequently, estimates of HIV outcomes that are obtained without addressing the underlying biases and challenges often encountered in these clinical databases will be less accurate. The general objective of the study was to use advanced statistical methods to improve estimates of HIV outcomes in the HIV clinical databases, and to use a nested research cohort database for validation.

Methods The study used data from a nested research cohort database from the Infectious Diseases Institute (IDI), Kampala. The clinical HIV databases used in the study included the IDI routine clinical database and the Academic Model for the Prevention and treatment of HIV/AIDS (AMPATH) database, in Eldoret. Statistical methods used included the following; calculation of a rate of under-reporting rate, imputation of missing data using chained equations, use of an inverse probability of weighting method by Frangakis & Rubin (F&R) and the nomogram method to correct mortality estimates for loss to follow-up, and finally the use of the inverse probability of censoring weights (IPCW) to account for patient drop-out.

Results The comparison of quality of data in the clinical database showed that there was a 1.8 – 2.3-fold higher rate of reporting of opportunistic events in the research cohort database compared to the IDI clinical database. The first year mortality estimates obtained from the IDI clinical database were greatly improved from 5.5% (95%CI 4.9% - 6.3%) to 11.2% (95% CI 5.8% -21.2%) using the F&R method, and to 11.9% (95% CI 8.0% - 15.7%) using the nomogram method. The corrected estimates were comparable to those obtained from the research cohort. The use of IPCW to adjust for patient drop out showed that during the two years after ART initiation, the differences between adjusted and unadjusted estimates diverged from 30 cells/μL, (assuming similar mortality and treatment access among dropouts as patients in care), to over 100 cells/μL (assuming 20% higher mortality and 50% lower treatment access among dropouts).

Conclusion Generally, ignoring the quality of data and naively analysing data from routine observational databases, without consideration of inherent biases results in significant under or over estimation of HIV outcomes response for patients starting treatment. Such results may overestimate the efficacy of HIV treatment programs. It is therefore important to utilize some of the recently developed statistical methodology to improve outcomes from routine HIV clinical databases in order to achieve substantial impact of quality of HIV care in Sub-Saharan Africa.