Adherence to Antiretroviral Therapy (ART) in the
DART Trial in Uganda and Zimbabwe
Statistical Analysis for Predictors and Consequences of Poor Adherence
Acta Universitatis Tamperensis No. 1748
By Sylvia Kiwuwa Muyingo
Tampere University Press
$87.50 Paper original
The aim of this doctoral thesis was to explore existing statistical methods and develop new tools to analyse adherence data. In addition to the development and description of statistical methods, this research tries to find answers to several important epidemiological questions. Analysis and understanding of adherence data is a big challenge for investigators and researchers. Poor medication adherence, for example, can lead to under-reporting of both therapeutic and adverse effects and undermine the results of the otherwise well-designed studies. In some clinical trials, optimal adherence cannot often be reached, and therefore adherence has a dual role in data analysis as an outcome and an important explanatory variable.
In this work, we analyse the data from a large cohort (n = 3316) of previously untreated African individuals initiating ART in rural and urban centres in Uganda and Zimbabwe. Participants were randomly assigned to receive laboratory and clinical monitoring (LCM), or clinically driven monitoring (CDM). We observed excellent clinic attendance over the first year on antiretroviral therapy (ART). Our follow-up included 93% of those enrolled. Adherence measured by drug possession ratio (DPR) was high at each visit. Only 12% of patients maintained consistently high adherence over the course of the first year. Most patients had high adherence most of the time, with only one or two visits with less than 95% adherence, and less than 1% of the participants never achieved high adherence during the first year. Regardless of the measure, adherence increases over the first year.
In this work we first explore different methods of summarising adherence data collected over a time interval. We consider traditional averaging approaches and quantile based classifications or groups of patients based on these. We also consider adherence data as a realization of a Markov chain, and use the estimated transition probabilities calculated separately for each individual as summary measures. Hierarchical clustering using these summary statistics is then used to classify the patients. Different classifications are compared by their interpretations and by cross-tabulations, the associations between group memberships and the relevant background variables are described, and the group memberships are used to predict the mortality and CD4 failures.
Generalized estimating equations (GEE) were used to model for optimal adherence during the first 48 weeks (12 visits). The impact of adherence during the first 48 weeks separately on time to death and time to CD4 failure was modeled with Cox proportional hazard models. Four different adherence classifications were used as explaining factors, and comparisons were made between the models. Finally, a dynamic logistic model was used to study the association between adherence and mortality. The model allows that the probability of dying between two clinic visits is explained by recent adherence history before the latest visit (assessed again at scheduled 4-weekly clinic visits) as well as by other (time dependent or baseline) covariates. In addition to the estimates of effects at the individual level, the approach also allows for the estimation of the population attributable fraction (PAF) a population level measure of the effect of adherence on mortality.
Based on our findings, a group of individuals (those with low CD4, reporting sexual partners 3 months prior to ART initiation, and low education) could be targeted for adherence-enhancing interventions both at ART initiation and in those not adhering well after a year on ART.
Worst adherence class based on Markov chain (MC) approach seems to predict mortality and CD4 failure independently of the worst class based on drug possession ratio (DPR).Whilst MC modeling is best suited to a research setting, DPR can be directly calculated from late return to clinic and self reports of 4-day/weekend a simple (does not require calculation) measure are therefore most suited to a clinical setting.
The estimated population attributable fractions (PAF) based on the dynamic logistic regression model, that is, the estimated proportions of deaths that could have been avoided with optimal adherence in the LCM and CDM groups during the 5 years follow-up period were 16.0% (90% CI -0.7,31.6)) and 33.1% (20.5,44.8), respectively. The estimated proportions of deaths on long-term ART that could be delayed at a population level (by eliminating non-optimal adherence) are similar to benefits from CD4 cell count monitoring of ART. In the absence of CD4 or viral load monitoring, individuals with optimal adherence experienced similar survival to those with customary adherence with CD4 monitoring suggesting that an alternative potential role of CD4 monitoring would be to reinforce adherence.
Return to Coronet Books main page